DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . This action is made final.
This Office Action is in response to the amendments filed of September 17th, 2025.
Claims 1-20 are pending in the case, claims 1, 19, and 20 have been amended.
Response to Amendment
The amendment filed on September 17th, 2025 has been entered. Claims 1-20 remain pending in the application.
Response to Arguments
Applicant's arguments filed September 17th, 2025 have been fully considered but they are not persuasive.
Regarding the 101 Arguments
Applicant argues:
Claims 1-20 is/are rejected under 35 U.S.C. § 101 because the claims are allegedly directed to an abstract idea without significantly more.
In the Office action, the Examiner states that:
Step 2A, Prong I
Claim 1 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., "machine learning model" [see MPEP 2106.04(a)(2)(III)]
"identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters" (e.g., a human can parse through recent data tables and determine if said data
deviates from a predetermined variable)
Accordingly, at Step 2A prong one, the claim is directed to an abstract idea.
Applicant respectfully traverse the Section 101 rejection in view of amended independent claims. The Applicant respectfully submits that amended independent claims 1, 19 and 20 are not directed to the alleged abstract idea, rather the claims are directed to smart collaborative unlearning of a machine learning model which is completely technical in nature and cannot be performed in the human mind or with the aid of a pencil and paper.
Further, according to reminders on evaluating subject matter eligibility of claims under 35 U.S.C. 101 dated August 4, 2025, at page 5.
III. Reminders on Deciding Whether to Make an SME Rejection
Examiners are reminded that if it is a “close call” as to whether a claim is eligible, they should only make a rejection when it is more likely than not (i.e. more than 50%) that the claim is ineligible under 35 U.S.C. 101. A Rejection of a claim should not be made simply because an examiner is uncertain as to the claim's eligibility.
In the Office action, Examiner contends only portion of claim such as step of identifying a most recent training iteration, comes under step 2A prong 1, that covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., "machine learning model") ... and Accordingly, at Step 2A prong one, the claim is directed to an abstract idea.
However, based on the reminder guidelines, "Examiners are reminded that if it is a "close call" as to whether a claim is eligible, they should only make a rejection when it is more likely than not (i.e., more than 50%) that the claim is ineligible".
According to the above reminder guidelines of USPTO, Applicant respectfully submits that amended claimed features such as "receiving a request to remove a dataset owned by a client from a machine learning model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client; in response to receiving the request, retraining the machine learning model by : identifying a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold; updating parameters of the machine learning model; performing one or more subsequent iterations of training of the machine learning model, and outputting the trained machine learning model for production to the client" are series of concrete technical steps, which as a whole, do not direct to an abstract idea.
Notwithstanding the above remark, assuming arguendo, that the previously filed independent claims 1, 19, and 20 are directed to an abstract idea/judicial exception as the Office Action contends, Applicant states that the amended independent claim 1 integrates a judicial exception into practical application. As indicated in the 2019 PEG, that in Prong Two, examiners should evaluate whether the claim as a whole integrates the recited judicial exception into a practical application of the judicial exception. Id. A claim that integrates a judicial exception into a practical application will apply, rely on, or use the judicial exception in a manner that imposes a meaningful limit on the judicial exception, such that the claim is more than a drafting effort designed to monopolize the judicial exception. Id.
Further, in the office action, the Examiner states that:
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a "machine learning model" which is recited at a high- level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)). The Examiner notes that this is used throughout the claim limitations, and is rejected thusly for each claim which recites the same language.
Regarding the "receiving a request to remove a dataset owned by a client from a machine learning model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client" limitation, the additional element of collecting data is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of selecting a particular data source or type of data to be manipulated in the claimed process (see MPEP 2106.05(g)). In addition, it is recited to perform the function of "updating parameters of the machine learning model, comprising adding noise to machine learning model para meters for the most recent training iteration", and "performing one or more subsequent iterations of training of the machine learning model, wherein the machine learning model is initialized with the updated parameters and the subsequent iterations train the machine learning model on multiple datasets excluding the dataset owned by the client", which are also recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
The amended independent claim 1, recites:
A computer implemented method comprising:
receiving a request to remove a dataset owned by a client from a machine learning
model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client;
in response to receiving the request, retraining the machine learning model by:
identifying, from stored noise sensitivities of the client, a most recent training
iteration that produced a noise sensitivity that is below a predetermined threshold,
wherein the predetermined threshold is based on a noise standard deviation and
predefined target privacy parameters;
updating parameters of the machine learning model, by adding noise to machine
learning model parameters for the most recent training iteration; and
performing one or more subsequent iterations of training of the machine
learning model, wherein the machine learning model is initialized with the updated
parameters and the subsequent iterations to train the machine learning model on multiple
datasets excluding the dataset owned by the client; and
outputting the trained machine learning model for production to end-users or clients.
(Emphasis added)
Applicant submits that the above emphasized claimed technical features address the technical problem that a lot of specific information from the different training data sets is included in the machine learning model, which leads to a noise amplitude that is too big to leave information from the kept training datasets in the noised model. Therefore, retraining is almost identical to training from scratch. (see paragraph 0005 of the specification)
The emphasized claimed features as a whole solves the above technical problem by enabling the system to informatively and automatically select a best global model to perform retraining on. As a result, anonymity criterion is satisfied, and the new model is obtained with a minimum amount of retraining. Further, a system implementing the presently described techniques achieves improved operational efficiency. Applicant notes that using the present invention if a client requests the removal of their dataset from the machine learning model, the system can continue functioning without interruption during the retraining process and the system remains fully operational regardless of whether a client chooses to opt out or discontinue their data usage. This results in a significant reduction in system downtime. Further, the present invention identifies noise sensitivities using a most recent training iteration that produced a noise sensitivity that enables the system to select a best global model to perform retraining on, retraining is more targeted and redundant training steps are not repeated. Accordingly, required bandwidth is reduced and system communication costs are optimized and minimized. (See paragraphs [0004]-[0005] of the specification).
The specification in at least these paragraphs [0053]-[0065] of the specification shows a practical implementation and how to perform unlearning of a machine learning model.
The claimed features such as receiving a request to remove a dataset owned by a client from a machine learning model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client; in response to receiving the request, retraining the machine learning model by: identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold,
wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters; updating parameters of the machine learning model, by adding noise to machine learning model parameters for the most recent training iteration; and performing one or more subsequent iterations of training of the machine learning model, wherein the machine learning model is initialized with the updated parameters and the subsequent iterations to train the machine learning model on multiple datasets excluding the dataset owned by the client; and outputting the trained machine learning model for production to clients do not recite an insignificant extra solution activity rather these steps involves an additional limitation (or combination) provides an improvement in the subsequent learning process which is much shorter than the original learning process, because the updated global machine learning model parameters retains information of the kept clients. (See paragraph 0058 of the specification)
Furthermore, the claim recites "receiving a request to remove a dataset owned by a client from a machine learning model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client" as supported by disclosure at paragraphs [0053]-[0054] of the applicant's originally filed specification which describes during stage (A) of example process, the server receives a request from one of the clients to forget the client. The request can be received during learning or after the machine learning model has been output for production and used for inference on new data. During stage (B), the server removes the requesting client from the multiple available clients. During stage (C), the server searches the noise sensitivities of the requesting client stored in the noise sensitivity storage. Applicant submits that collection of data is not extra-solution activity as it tracks or manages the client-specific data and sensitivities which cannot be performed in mind or with aid of a pencil/paper, rather the claims physically integrate the judicial exception into a practical application and provide the improvement.
Further, according to reminders on evaluating subject matter eligibility of claims under 35 U.S.C. 101 dated August 4, 2025, at pages 3-4.
B. Step 2A Prong Two
Analysis of claim as a whole: The analysis in Step 2A Prong Two considers the claim as a whole. The way in which the additional elements use or interact may integrate the judicial exception into a practical application. Accordingly, the additional limitations should not be evaluated in a vacuum, completely separate from the recited judicial exception. Instead, the analysis should take into consideration all the claim limitations and how these limitations interact and impact each other when evaluating whether the exception is integrated into a practical application.
While an additional limitation (or combination) that merely applies the judicial exception on a generic computer may not render a claim eligible on its own, an additional limitation (or combination) that meaningfully limits
the judicial exception can render it eligible.
Improvements consideration: In computer-related technologies, examiners can conclude that claims are eligible in Step 2A Prong Two by finding that a claim reflects an improvement to the functioning of a computer or to another technology or technical field, integrating a recited judicial exception into a practical application of the exception. This consideration has also been referred to as the search for a technological solution to a technological problem. An important consideration in determining whether a claim improves technology or a technical field is the extent to which the claim covers a particular solution to a problem or a particular way to achieve a desired outcome, as opposed to merely claiming the idea of a solution or outcome. The examiner is reminded to consult the specification to determine whether the disclosed invention improves technology or a technical field, and evaluate the claim to ensure it reflects the disclosed improvement. The specification does not need to explicitly set forth the improvement, but it must describe the invention such that the improvement would be apparent to one of ordinary skill in the art. The claim itself does not need to explicitly recite the improvement described in the
specification.
In view of the foregoing remarks and above reminder guidelines from USPTO, the Applicant respectfully submits that amended independent claim 1 integrates the purported judicial exception into the practical application as the amended independent claim 1 clearly identifies that "claim improves technology or a technical field ." Further, the amended independent claim 1 "covers a particular solution to a problem or a particular way to achieve a desired outcome" as reminded in the guidelines.
Accordingly, the Applicant submits that claim 1, as amended, is not directed to the alleged abstract idea and is instead directed to patent eligible subject matter.
Independent claims 19 and 20, as amended, recite similar features. Therefore, claims 19 and 20 are also directed to patent eligible subject matter. Dependent claims 1-18 are also directed to patent eligible subject matter based on their dependency on their respective independent claim.
Accordingly, Applicant respectfully requests that the Examiner reconsider and withdraw the rejection of claims 1-20 under 35 U.S.C. § 101.
Examiners Response
Applicants’ arguments and amended claims still continue to recite the concept of determining a model update checkpoint by comparing computed noise sensitivities against a threshold, which constitutes a mental concept and data evaluation under MPEP 2106.04(a)(2)(III). The additional limitations of receiving a client request, retraining a model, and outputting for production merely recite generic computer implementation and post-solution activity. The claim does not specify any technical manner in which the computer or model is improved. Rather, it applies the abstract idea using routine machine learning operations, thereby failing to integrate the exception into a practical application. Accordingly, the claim remains directed to an abstract idea and does not recite significantly more, the rejection is maintained.
Regarding the 103 arguments
Applicant argues:
Claims 1-6, 10-17, 19, and 20 were rejected under 35 U.S.C. § 103 as allegedly being unpatentable over Bourtoule et al. ("Machine Unlearning", referred to as Bourtoule) in view of Guo et al. ("Certified Data Removal from Machine Learning Models", referred to as Guo).
Claims 7, 8, and 9 were rejected under 35 U.S.C. § 103 as allegedly being unpatentable over Bourtoule et al. ("Machine Unlearning", referred to as Bourtoule) in view of Guo et al. ("Certified Data Removal from Machine Learning Models", referred to as Guo) in view of Abadi et al. ("Deep Learning with Differential Privacy", referred to as Abadi.
Claim 18 was rejected under 35 U.S.C. § 103 as allegedly being unpatentable over Bourtoule et al. ("Machine Unlearning", referred to as Bourtoule) in view of Guo et al. ("Certified Data Removal from Machine Learning Models", referred to as Guo) in view of McMahan et al. ("Communication-Efficient Learning of Deep Networks from Decentralized Data", referred to as McMahan).
Independent claim 1 recites, inter alia
"identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters;"
Further, in the office action the Examiner contends and relies on Guo page 5 algorithm 2 for claim 1 such as:
"identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters."
(Page 5, Algorithm 2: Describes accumulating the clients noise sensitivity's into β for each iteration and then comparing [β to the DE-derived threshold
σ
ϵ
∕
c
. The algorithm only applies the approximate removal update when is less than or equal to
σ
ϵ
∕
c
, and retrains the prior iteration where β first exceeded
σ
ϵ
∕
c
, corresponding to a threshold.);
Applicant respectfully traverse and submits that Guo as relied upon on text at Page 5, Algorithm 2, describes repeated certified removal of data batches if gradient residual bound (β)> parameters (
σ
ϵ
∕
c
)
then re-train from scratch using Algorithm 1. Further, Guo at page 5, describes that During removal (line 19 in Algorithm 2), we apply the batch Newton update (Equation 6) and compute the gradient residual norm bound using Corollary 2 (line 15 in Algorithm 2). The variable p accumulates the gradient residual norm over all removals. If the pre-determined budget of
σ
ϵ
∕
c
is exceeded, we train a new removal-enabled model from scratch using Algorithm on the remaining data points.
It is evident from above that Guo merely teaches a new removal-enabled model is trained from scratch when pre-determined budget of 6E/c is exceeded. However, the present invention teaches that the server searches the noise sensitivities of the requested client stored in the noise sensitivity storage to identify a most recent training/learning iteration that produced a noise sensitivity for the requesting client that is below a predetermined threshold.
Accordingly, Guo fails to teach or suggest "identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters."
Further, Bourtoule, Abadi and McMahan do not remedy the deficiency of the Guo.
Accordingly, none of the references individually or in combination teach or suggest the above features.
Dependent claims 2-18 depending on independent claims 1, 19 and 20 are patentably distinct for at least the reasons discussed with respect to amended independent claims 1, 19 and 20.
Further, withdrawal of the rejections and allowance of claims 1-20 are therefore respectfully requested.
Therefore, for the above reasons the rejection under 35 U.S.C. § 103 is respectfully traversed and withdrawal of the rejection is respectfully requested.
Examiners response
Applicant’s arguments have been fully considered but is not persuasive. Guo et al. teaches tracking client-specific privacy and sensitivity metrics for each training round and using a threshold comparison to determine whether to retrain or resume form a prior iteration (Algorithm 2, p. 5). The act of evaluating the accumulated sensitivity for each iteration and determining whether it remains below the privacy budget corresponds to identifying a most recent iteration at which the noise sensitivity is less than a predetermined threshold. Guo performs the claimed step implicitly by maintaining the per-iteration sensitivity history and comparing it to a threshold derived from the parameters. The selection of the iteration at which β<
σ
ϵ
∕
c
corresponds to the claimed “most recent training iteration”. Bourtoule further teaches the overall unlearning framework that removes a specific client dataset and retrains the model accordingly. Guo discloses evaluating β relative to
σ
ϵ
∕
c
before deciding to retrain necessarily entails identifying the most recent iteration. Accordingly, the combination of Bourtoule, Guo and Abadi still stands and the rejection is maintained.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
To determine if a claim is directed to patent ineligible subject matter, the Court has guided the Office to apply the Alice/Mayo test, which requires:
Step 1: Determining if the claim falls within a statutory category.
Step 2A: Determining if the claim is directed to a patent ineligible judicial exception consisting of a law of nature, a natural phenomenon, or abstract idea; and Step 2A is a two prong inquiry. MPEP 2106.04(II)(A). Under the first prong, examiners evaluate whether a law of nature, natural phenomenon, or abstract idea is set forth or described in the claim. Abstract ideas include mathematical concepts, certain methods of organizing human activity, and mental processes. MPEP 2104.04(a)(2). The second prong is an inquiry into whether the claim integrates a judicial exception into a practical application. MPEP 2106.04(d).
Step 2B: If the claim is directed to a judicial exception, determining if the claim recites limitations or elements that amount to significantly more than the judicial exception. (See MPEP 2106).
Claims 1-20 is/are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1-18 are directed to a method (a process), and Claim 19 is directed to a system (a machine). Therefore, Claims 1-19 are directed to a process, machine or manufacture or composition of matter.
Claim 20, under its broadest reasonable interpretation, is directed to a computer readable storage medium, which encompasses transitory, propagating signals (see [0101] of applicant’s specification). Transitory signals are not a process, machine or manufacture or composition of matter. Accordingly, the claim is directed to non-statutory subject matter and is rejected under 35 U.S.C § 101 (See MPEP 2106.03(II)).
Regarding claim 1
Step 2A, Prong 1
Claim 1 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters” (e.g., a human can parse through recent data tables and determine if said data deviates from a predetermined variable)
Accordingly, at Step 2A, prong one, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “machine learning model” which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)). The Examiner notes that this is used throughout the claim limitations, and is rejected thusly for each claim which recites the same language.
Regarding the “receiving a request to remove a dataset owned by a client from a machine learning model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client” limitation, the additional element of collecting data is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of selecting a particular data source or type of data to be manipulated in the claimed process (see MPEP 2106.05(g)). In addition, it is recited to perform the function of “in response to receiving the request, retraining the machine learning model”, “updating parameters of the machine learning model, by comprising adding noise to machine learning model parameters for the most recent training iteration”, and “performing one or more subsequent iterations of training of the machine learning model, wherein the machine learning model is initialized with the updated parameters and the subsequent iterations to train the machine learning model on multiple datasets excluding the dataset owned by the client”, which are also recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “outputting the trained machine learning model for production to the client”, limitation, this additional element is recited at a high-level of generality and amounts to extra-solution activity of outputting data to from a model, i.e., post-solution activity of data outputting (see MPEP 2106.05(g)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of a “machine learning model” which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “receiving a request to remove a dataset owned by a client from a machine learning model, the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client” limitation, as discussed above, the additional element of collecting data is recited at a high level of generality and amounts to extra-solution activity of receiving data i.e. pre-solution activity of selecting a particular data source or type of data to be manipulated in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory") which is also recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “in response to receiving the request, retraining the machine learning model”, “updating parameters of the machine learning model, by comprising adding noise to machine learning model parameters for the most recent training iteration”, and “performing one or more subsequent iterations of training of the machine learning model, wherein the machine learning model is initialized with the updated parameters and the subsequent iterations to train the machine learning model on multiple datasets excluding the dataset owned by the client”, limitations, which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “outputting the trained machine learning model for production to the client”, limitation, as discussed above, the additional element of outputting data is recited at a high level of generality and amounts to extra-solution activity of outputting data for the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II),
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 2
Step 2A, Prong 1
Claim 2 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“wherein each noise sensitivity bounds a difference between the machine learning model trained on the multiple datasets and the machine learning model trained on multiple datasets excluding the dataset owned by the client” (e.g., a comparison of two sets of model values)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception
Regarding claim 3
Step 2A, Prong 1
Claim 3 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “wherein the noise sensitivities track respective evolutions of an amount of client information included in the machine learning model during training” the additional element of collecting data is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. post-solution activity of gathering data for use in the claimed process (see MPEP 2106.05(g)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “wherein the noise sensitivities track respective evolutions of an amount of client information included in the machine learning model during training” is merely collecting data and is recited at a high level of generality and amounts to extra-solution activity of receiving data i.e. post-solution activity of gathering data for use in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 4
Step 2A, Prong 1
Claim 4 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“wherein the noise sensitivity of the client is based on a difference between i) an aggregation of machine learning model parameters over the multiple datasets and ii) an aggregation of machine learning model parameters over the multiple datasets excluding the dataset owned by the client” (e.g., a comparison of two sets of model values)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception
Regarding claim 5
Step 2A, Prong 1
Claim 5 recites the following mathematical concepts, that in each case under the broadest reasonable interpretation, covers performance of mathematical relationships, mathematical formulas or equations, and mathematical calculations but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(I)].
“wherein the noise sensitivity of the client comprises a sum, over all preceding training iterations, of a Euclidean norm of the differences” (e.g., a summing and computing a norm from sets of model values)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception
Regarding claim 6
Step 2A, Prong 1
Claim 6 recites the following mathematical concepts, that in each case under the broadest reasonable interpretation, covers performance of mathematical relationships, mathematical formulas or equations, and mathematical calculations but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(I)].
“wherein the noise sensitivity of the client comprises a weighted sum, over all preceding training iterations, of a Euclidean norm of the differences, wherein the weights are based on regularization and convexity parameters of loss functions used by the clients” (e.g., a summing, computing a norm, and weight calculations from sets of model values)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 7
Step 2A, Prong 1
Claim 7 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “wherein the noise standard deviation is common to each client” the additional element is merely recited to describe parameter settings for the mathematical concepts.
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “wherein the noise standard deviation is common to each client” the additional element is describing the implementation of the abstract idea.
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 8
Step 2A, Prong 1
Claim 8 recites the following mathematical concepts, that in each case under the broadest reasonable interpretation, covers performance of mathematical relationships, mathematical formulas or equations, and mathematical calculations but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(I)].
“wherein the predetermined threshold is given by ca/c where E represents a first predefined target privacy parameter, a represents the noise standard deviation, and c= 2 ln(1.25/) where 6 represents a second predefined target privacy parameter” (e.g., a arithmetic formula and their assigned symbols)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception
Regarding claim 9
Step 2A, Prong 1
Claim 9 recites the following mathematical concepts, that in each case under the broadest reasonable interpretation, covers performance of mathematical relationships, mathematical formulas or equations, and mathematical calculations but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(I)].
“wherein adding noise to machine learning model parameters comprises adding noise sampled from a normal distribution with zero mean” (e.g., a statistical act of drawing random values from a Gassian distribution)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 10
Step 2A, Prong 1
Claim 10 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “receiving a series of requests, each request requesting removal of one or more datasets owned by a respective set of clients from the machine learning model” limitation, the additional element is recited at a high-level of generality and amounts to extra-solution activity of retrieving requests of instructions, i.e., pre-solution activity of data gathering for use in the claimed system (see MPEP 2106.05(g)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “receiving a series of requests, each request requesting removal of one or more datasets owned by a respective set of clients from the machine learning model”, limitation, the additional element is recited at a high-level of generality and amounts to extra-solution activity of retrieving updated input values from network sources, i.e. pre-solution activity of data gathering. The courts have found limitations directed to obtaining information electronically, recited at a high-level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 11
Step 2A, Prong 1
Claim 11 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“wherein the noise sensitivities determined during training of the machine learning model comprise a maximum noise sensitivity from a set of noise sensitivities computed for the set of clients” (e.g., a difference between values and determining a maximum value form the set of values)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 12
Step 2A, Prong 1
Claim 12 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“wherein each noise sensitivity in the set of noise sensitivities comprises a difference between i) an aggregation of machine learning model parameters over a federated dataset of remaining clients after forgetting a client in the set of clients and ii) an aggregation of machine learning model parameters over the federated dataset of remaining clients after forgetting clients in a same request excluding a dataset owned by the client in the set of clients” (e.g., a difference between values by staking in numeric summaries and subtracting one form the other)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 13
Step 2A, Prong 1
Claim 13 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“wherein the most recent training iteration that produced a noise sensitivity that is below a predetermined threshold is dependent on an index that represents a first training iteration for which noise sensitivities of the set of clients is above the predetermined threshold” (e.g., a logical rule for choosing one numeric iteration index over another based on a comparison, by selecting a value compared to a threshold)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 14
Step 2A, Prong 1
Claim 14 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“removing the client from a list of currently available clients” (e.g.,)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 15
Step 2A, Prong 1
Claim 15 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “wherein the request to remove the dataset comprises a request to remove a dataset owned by one or more of: an attacker, a client that induces bias in the machine learning model, or a client that does not respect general data protection regulations”, limitation, which is recited at a high-level of generality such that it amounts to a description of who generates the request, which is a field of use and technological environment (See MPEP 2106.05(h)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “wherein the request to remove the dataset comprises a request to remove a dataset owned by one or more of: an attacker, a client that induces bias in the machine learning model, or a client that does not respect general data protection regulations”, limitation, which is recited at a high-level of generality such that it amounts to a description of a field of use and technological environment (See MPEP 2106.05(h)).
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 16
Step 2A, Prong 1
Claim 16 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
“” (e.g., a comparison of two values)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of “wherein a number of subsequent iterations of training is less than a number of previous iterations performed during previous training of the machine learning model” limitation, this additional element is recited at a high level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “wherein a number of subsequent iterations of training is less than a number of previous iterations performed during previous training of the machine learning model” which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2B, the additional elements individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 17
Step 2A, Prong 1
Claim 17 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “using the trained machine learning model for inference”, limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “using the trained machine learning model for inference”, limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 18
Step 2A, Prong 1
Claim 18 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “machine learning model”) [see MPEP 2106.04(a)(2)(III)].
“aggregating the received machine learning model parameters for the iteration” (e.g., an averaging returned value for further training)
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional element of a “initializing the machine learning model on initial machine learning model parameters”, limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)). In addition, this is merely recited in order to perform the function “for each of multiple iterations” and does not render the claim subject matter eligible.
Regarding the “providing the clients with the initial machine learning model parameters or machine learning model parameters for a previous iteration, wherein each client uses a respective dataset to update the initial machine learning model parameters or machine learning model parameters for the previous iteration”, and “providing the aggregated machine learning model parameters for the iteration as input for a subsequent iteration” limitations, the additional elements are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “receiving, from each of the clients, machine learning model parameters for the iteration” limitation, the additional element of collecting data is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. post-solution activity of gathering data for use in the claimed process (see MPEP 2106.05(g)).
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “using the trained machine learning model for inference”, limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “providing the clients with the initial machine learning model parameters or machine learning model parameters for a previous iteration, wherein each client uses a respective dataset to update the initial machine learning model parameters or machine learning model parameters for the previous iteration”, and “providing the aggregated machine learning model parameters for the iteration as input for a subsequent iteration” limitations, as discussed above, the additional elements are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “receiving, from each of the clients, machine learning model parameters for the iteration” limitation, as discussed above, the additional element of collecting data is recited at a high level of generality and amounts to extra-solution activity of receiving data i.e. post-solution activity of gathering data for use in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 19
Claim 19 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional elements of “ a system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations” limitation, the additional elements are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)). The claim corresponds directly to the method steps of claim 1, respectively, with the addition of generic hardware components which are insufficient to render the claims subject matter eligible for the same reasons described above.
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “ a system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations” limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 20
Claim 20 does not introduce any new abstract ideas, but is directed to the abstract idea identified in claim 1.
Accordingly, at Step 2A, prong 1, the claim is directed to an abstract idea.
Step 2A, Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional elements of “ non-transitory computer-readable storage medium comprising instructions stored thereon that are executable by a processing device and upon such execution cause the processing device to perform operations” limitation, the additional elements are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)). The claim corresponds directly to the method steps of claim 1, respectively, with the addition of generic hardware components which are insufficient to render the claims subject matter eligible for the same reasons described above.
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of “computer-readable storage medium comprising instructions stored thereon that are executable by a processing device and upon such execution cause the processing device to perform operations” limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-6, 10-17, 19, and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bourtoule et al. ("Machine Unlearning", referred to as Bourtoule) in view of Guo et al. ("Certified Data Removal from Machine Learning Models", referred to as Guo).
Regard claim 1, Bourtoule teaches, a computer implemented method comprising ([):
receiving a request to remove a dataset owned by a client from a machine learning model(Page 7, ¶2: Describes a unlearning request API that takes a clients data slice ID and triggers the rollback procedure or that slice.),
performing one or more subsequent iterations of training of the machine learning model, wherein the machine learning model is initialized with the updated parameters and the subsequent iterations to train the machine learning model on multiple datasets excluding the dataset owned by the client (Page 6, Fig. 2, ¶1: “When a data point is to be unlearned, only the constituent model whose dataset contains this point is affected. More specifically, a data point is unlearned from a particular slice in a particular shard. Retraining can start from the last parameter state saved prior to including the slice containing the data point to be unlearned:...” Describes that upon removal of a clients data point, SISA training reinitializes the affected model at the last saved checkpoint/updated parameters and recontinues training on the remaining data slices, excluding the removed points slice.).
outputting the trained machine learning model for production to the client (Page 5, Section IV :Describes providing the trained/unlearned model for inference/production use by aggregation constituent models’ predictions and outputting a final result to be consumed, corresponding to an outputted model for a client.).
Although Bourtoule teaches receiving a request to remove a dataset owned by a client from a machine learning model and performing one or more subsequent iterations of training of the machine learning model, wherein the machine learning model is initialized with the updated parameters and the subsequent iterations train the machine learning model on multiple datasets excluding the dataset owned by the client, and outputting the machine learning model for production to the client, it does not teach the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client, in response to receiving the request identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters and updating parameters of the machine learning model, comprising adding noise to machine learning model parameters for the most recent training iteration.
Guo teaches the machine learning model being associated with a set of noise sensitivities determined during training of the machine learning model on multiple datasets owned by respective clients including the client (Page 5. ¶3: “The variable β accumulates the gradient residual norm over all removals. If the pre-determined budget of σϵ/c is exceeded, we train a new removal-enabled model from scratch using Algorithm 1 on the remaining data points.” Describes a store and update per-client gradient-residual norm β, during training and removal. The associated σare tied to each β which corresponds to noise sensitivity metric.); and
in response to receiving the request, retraining the machine learning model(Page 5 Algorithm 2: Shows that the model is run in a for loop, indicating that it is retrained as necessary and re-trains from scratch using algorithm 1.) by:
identifying, from stored noise sensitivities of the client, a most recent training iteration that produced a noise sensitivity that is below a predetermined threshold, wherein the predetermined threshold is based on a noise standard deviation and predefined target privacy parameters. (Page 5, Algorithm 2: Describes accumulating the clients noise sensitivity’s into β for each iteration and then comparing β to the DE-derived threshold σϵ/c. The algorithm only applies the approximate removal update when is less than or equal to σϵ/c, and retrains the prior iteration where β first exceeded σϵ/c, corresponding to a threshold.);
updating parameters of the machine learning model, by adding noise to machine learning model parameters for the most recent training iteration (Page 5, Algorithm 1: Describes a perturb training loss by sampling a Gaussian vector b of variances σ2 and adding the linear term btw to the objective. This has the effect of adding noise to the learned parameters at the specified iteration.)
It would have been obvious to a person of ordinary skill in the art at the time of the claimed invention to combine the SISA unlearning framework of Bourtoule with the certified removal mechanism of Guo. Doing so would allow for more efficient retraining of models to by identifying data to properly and securely remove upon requests received.
Regarding claim 2, Guo further teaches, wherein each noise sensitivity bounds a difference between the machine learning model trained on the multiple datasets and the machine learning model trained on multiple datasets excluding the dataset owned by the client(Page 5, Section 3.2, Describes removing data from models using multiple thermos, then after this has been applied to the training model, it is verified, and compared, if data still exists that is needed to be removed still, it will create a new removal enabled model for the remaining data points, this corresponds to comparing models with previous models having data removed for a client.).
Regarding claim 3, Guo further teaches, wherein the noise sensitivities track respective evolutions of an amount of client information included in the machine learning model during training (Page 5, Algorithm 2: Describes how β is updated each removal by adding the new gradient residual norm, so βalways reflects the information retained of that client’s data on the model. Β tracks the evolution of how much client specific information remains embedded in the model as training proceeds.) .
Regarding claim 4, Guo further teaches, wherein the noise sensitivity of the client is based on a difference between i) an aggregation of machine learning model parameters over the multiple datasets and ii) an aggregation of machine learning model parameters over the multiple datasets excluding the dataset owned by the client. (Page 2, Section 3, ¶3 Removal mechanism: Describes Δ that captures the difference between the full data solution w* (aggregation over all datasets) and the leave-one-out solution for client B( aggregation excluding that client’s points. Then in theorem 1, describes the noise sensitivity, which corresponds to matching metrics based on parameter difference between two aggregations.)
Regarding claim 5, Guo further teaches, wherein the noise sensitivity of the client comprises a sum, over all preceding training iterations, of a Euclidean norm of the differences. (Page 5, Algorithm 2: Describes that each iteration adds a term proportional to the Euclidean norm ∥H−1Δ∥2(leave-one-out parameter difference) into β. Over j removals, βbecomes the sum of these per-iteration norms.)
Regarding claim 6, Guo further teaches, wherein the noise sensitivity of the client comprises a weighted sum, over all preceding training iterations, of a Euclidean norm of the differences, wherein the weights are based on regularization and convexity parameters of loss functions used by the clients. (Page 5, Algorithm 2: Describes each term ∥H−1Δ∥2 (core Euclidean norm) is weighted by γ(convexity/Lipschitz parameter) and by K∥2 ・ ∥H−1Δ∥2 , data dependent norms reflecting regularization and Hessian structure. These yield a weighted sum of Euclidean norms across iterations.)
Regarding claim 10, Bourtoule further teaches, wherein the method further comprises receiving a series of requests, each request requesting removal of one or more datasets owned by a respective set of clients from the machine learning model(Page 6, Fig. 2 ¶1: Describes a framework that handles multiple unlearning requests over time, each new request triggers the same slice based rollback for the client(s) named in the request.).
Regarding claim 11, Guo further teaches, wherein the noise sensitivities determined during training of the machine learning model comprise a maximum noise sensitivity from a set of noise sensitivities computed for the set of clients(Page 5, ¶2 Pseudo-code: Describes the removal budget as a maximum over all per-client gradient-residual norms (noise sensitivities). The budget is tied to σϵ/c, which are selected from the highest sensitivity from the set computed for all clients.).
Regarding claim 12, Guo further teaches, wherein each noise sensitivity in the set of noise sensitivities comprises a difference between i) an aggregation of machine learning model parameters over a federated dataset of remaining clients after forgetting a client in the set of clients and ii) an aggregation of machine learning model parameters over the federated dataset of remaining clients after forgetting clients in a same request excluding a dataset owned by the client in the set of clients(Page 2, ¶5, Removal Mechanism: Describes w* as the model parameter vector trained on the full dataset(aggregation over all clients), and w- is the leave-out-batch solution(aggregation excluding that clients data). The vector Δ corresponds as the difference between those two aggregations.).
Regarding claim 13, Guo further teaches, wherein the most recent training iteration that produced a noise sensitivity that is below a predetermined threshold is dependent on an index that represents a first training iteration for which noise sensitivities of the set of clients is above the predetermined threshold(Page 5, Algorithm 2: Describes that the selection of the rollback point is dependent on the iteration index at which the threshold is first crossed.).
Regarding claim 14, Bourtoule further teaches, wherein the method further comprises, in response to receiving the request, removing the client from a list of currently available clients(Page 6, Fig. 2, ¶1: Describes that SISA training, each client’s data resides in its own “slice”. Upon receiving an unlearning request, that slice is removed (the client is dropped from the set of active slices), and retraining proceeds without it.).
Regarding claim 15, Bourtoule further teaches, wherein the request to remove the dataset comprises a request to remove a dataset owned by one or more of: an attacker, a client that induces bias in the machine learning model, or a client that does not respect general data protection regulations(Page 1, Abstract: Describes that the unlearning framework is designed to handle any data-deletion request, which encompasses deletion due to attack, bias or general data protection regulations compliance.).
Regarding claim 16, Guo further teaches, wherein a number of subsequent iterations of training is less than a number of previous iterations performed during previous training of the machine learning model(Page 5, Algorithm 2: Describes that when β > σϵ/c the algorithm does not invoke the multi-epoch retraining but instead executes exactly one parameter update (w ← w + H−1Δ) the number of subsequent update iterations post-removal is just one, which is fewer that the original number of epochs used during the original training.).
Regarding claim 17, Bourtoule further teaches, further comprising using the trained machine learning model for inference (Page 7, Section 4 Aggregation: “inference time, predictions from various constituent models can be used to provide an overall prediction. The choice of aggregation strategy in SISA training is influenced by two key factors” describes that the SISA training uses inference in its repeated training of data removal processes.).
Regarding claim 19, which recite substantially the same limitations as claim 1. Claim 19 further recites, a system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations (Bortoule, Page 1, Abstract and Section 1: Describes data stored online from/on computers and their implementation of unlearning data from learned models with the described steps. The examiner notes that the learned machine learning models sharing data online and offline to be unlearned utilizing computer components inherently requires, hardware comprised of, but not limited to, processors, RAM, hard drives, servers, etc.) to perform the method steps of claim 1, respectively, and is therefore rejected on the same premise.
Regarding claim 20, which recite substantially the same limitations as claim 1. Claim 20 further recites, a non-transitory computer-readable storage medium comprising instructions stored thereon that are executable by a processing device and upon such execution cause the processing device to perform operations (Bortoule, Page 1, Abstract and Section 1: Describes data stored online from/on computers and their implementation of unlearning data from learned models with the described steps. The examiner notes that the learned machine learning models sharing data online and offline to be unlearned utilizing computer components inherently requires, hardware comprised of, but not limited to, processors, RAM, hard drives, servers, etc.) to perform the method steps of claim 1, respectively, and is therefore rejected on the same premise.
Claim(s) 7, 8, and 9 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bourtoule et al. ("Machine Unlearning", referred to as Bourtoule) in view of Guo et al. ("Certified Data Removal from Machine Learning Models", referred to as Guo) in view of Abadi et al. ("Deep Learning with Differential Privacy", referred to as Abadi.
Regarding claim 7, Abadi teaches, wherein the noise standard deviation is common to each client(Page 3, Algorithm 1: Describes taking in a single, global noise scale σ as an input parameter that is applied identically in every gradient step and for every client’s contribution, σ is common across clients. ).
Regarding claim 8. Abadi further teaches wherein the predetermined threshold is given by ca/c where E represents a first predefined target privacy parameter, a represents the noise standard deviation, and c= 2 ln(1.25/) where 6 represents a second predefined target privacy parameter(Page 3, ¶8 Moments accountant: Describes that to gather (ε,δ)-differential privacy the Gaussian noise scale σ must be chosen so that
σ
≥
2
\
l
n
1.25
δ
ε
, which rearranges to the exact “ε σ/c” threshold.).
Regarding claim 9. Abadi further teaches wherein adding noise to machine learning model parameters comprises adding noise sampled from a normal distribution with zero mean(Page 3, Algorithm 1: Describes a zero-mean Gaussian noise which is used to perturb the clipped gradient.).
It would have been obvious to a person of ordinary skill in the art at the time of the claimed invention to combine the SISA unlearning framework of Bourtoule , the certified removal mechanism of Guo with the noise calibration of Abadi. Doing so would provide uniform noise standard deviation to ensure efficient retraining of models for streamlined integration.
Claim(s) 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bourtoule et al. ("Machine Unlearning", referred to as Bourtoule) in view of Guo et al. ("Certified Data Removal from Machine Learning Models", referred to as Guo) in view of McMahan et al. ("Communication-Efficient Learning of Deep Networks from Decentralized Data", referred to as McMahan).
Regarding claim 18, McMahan teaches, wherein training the machine learning model on the multiple datasets comprises:
initializing the machine learning model on initial machine learning model parameters (Page 4, ¶2: Describes that a command is used to initialize w0 which shows that the server begins the federated procedure with a concrete shared model parameter vector.) ; for each of multiple iterations:
providing the clients with the initial machine learning model parameters or machine learning model parameters for a previous iteration, wherein each client uses a respective dataset to update the initial machine learning model parameters or machine learning model parameters for the previous iteration (Page 2, ¶7: Describes that the server sends the current global algorithm state and that each client performs local computations on that state using its own data, FedAvg teaches both broadcast of model parameters and the per-client local update using the client’s dataset.);
receiving, from each of the clients, machine learning model parameters for the iteration (Page 2, ¶7: Describes that it sends an update to the server, which, after local training, each client returns its updated model parameters back to the central server.);
aggregating the received machine learning model parameters for the iteration (Page 2, ¶7: Describes a weighted-average formula which aggregates model parameters, and the server combines the clients returned parameters into the next global model.) ; and providing the aggregated machine learning model parameters for the iteration as input for a subsequent iteration (Page 2, ¶7: Describes that after the aggregation the process repeats, FedAvg takes the newly computed Wt+1 and feds it back into the next round.).
It would have been obvious to a person of ordinary skill in the art at the time of the claimed invention to combine the SISA unlearning framework of Bourtoule , the certified removal mechanism of Guo with the Federated Averaging protocol of McMahan. Doing so would continuously retain models for unlearning by incorporating standard federated training for efficient model updates throughout its steps.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DONALD T RODEN whose telephone number is (571)272-6441. The examiner can normally be reached Mon-Thur 8:00-5:00 EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Omar Fernandez Rivas can be reached at (571) 272-2589. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/D.T.R./Examiner, Art Unit 2128
/OMAR F FERNANDEZ RIVAS/Supervisory Patent Examiner, Art Unit 2128
Prosecution Insights