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 .
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.
Step 1
According to the first part of the analysis, in the instant case, claims 1-8, 17-20 are directed to a method, and claims 9-17 are directed to an apparatus. Each of these claims fall within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter).
For claim 1,
Step 2A Prong One
selecting clients from the set of clients based on the prediction confidence values included in the updates;
(Selecting is a mental process)
Step 2A Prong Two
receiving an update from each client in a set of clients of a federated learning system for a current training round
(Receiving updates [data gathering] is insignificant extra-solution activity. See MPEP § 2106.05(g))
where each of the clients trained a local model using local data,
(Generic model training is insignificant extra-solution activity. See MPEP § 2106.05(g))
and sending the new global model to each of the clients in the set of clients.
(Sending the new global model [transmitting data] is insignificant extra-solution activity. See MPEP § 2106.05(g))
wherein each of the updates includes a prediction confidence value and a model update;… generating a new global model using the model updates in the updates received from the selected clients;
These additional elements integrate the claim as a whole into a practical application. Thus, claim 1 is not rejected under 35 U.S.C. 101.
For claim 9,
Step 2A Prong One
selecting clients from the set of clients based on the prediction confidence values included in the updates;
(Selecting is a mental process)
Step 2A Prong Two
A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform operations comprising:
(Implementing on a generic computer is mere instructions to apply a judicial exception. See MPEP § 2106.05(f))
receiving an update from each client in a set of clients of a federated learning system for a current training round
(Receiving updates [data gathering] is insignificant extra-solution activity. See MPEP § 2106.05(g))
where each of the clients trained a local model using local data,
(Generic model training is insignificant extra-solution activity. See MPEP § 2106.05(g))
and sending the new global model to each of the clients in the set of clients.
(Sending the new global model [transmitting data] is insignificant extra-solution activity. See MPEP § 2106.05(g))
wherein each of the updates includes a prediction confidence value and a model update; … generating a new global model using the model updates in the updates received from the selected clients;
These additional elements integrate the claims a whole into a practical application. Thus, claim 9 is not rejected under 35 U.S.C. 101.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claim(s) 1, 8-9, 16 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Zhuowei Wang et al. (hereinafter Wang) (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022).
Regarding claim 1, Wang teaches;
A method comprising: receiving an update from each client in a set of clients of a federated learning system for a current training round
[pg. 4]
PNG
media_image1.png
36
549
media_image1.png
Greyscale
PNG
media_image2.png
19
256
media_image2.png
Greyscale
PNG
media_image3.png
312
634
media_image3.png
Greyscale
[pg. 3]
PNG
media_image4.png
125
575
media_image4.png
Greyscale
NOTE: Wang teaches that the central server receives updates θ_k from each client k in the set of clients S of a federated learning system for a current training round r.
where each of the clients trained a local model using local data,
[pg. 3]
PNG
media_image5.png
113
1044
media_image5.png
Greyscale
wherein each of the updates includes a prediction confidence value and a model update;
([pg. 4] Given the confidence scores for all samples in Dk, we can estimate the number of clean labels on client-k as sk by summing up the confidence scores of all the samples on the client. Each client can compute the score sk using its received global model and send sk back to the server.)
[pg. 3]
PNG
media_image4.png
125
575
media_image4.png
Greyscale
NOTE: Wang teaches prediction confidence s_k and the model update θ_k for each client k which are sent to the server to update the global model.
selecting clients from the set of clients based on the prediction confidence values included in the updates;
[pg. 4]
PNG
media_image6.png
26
583
media_image6.png
Greyscale
NOTE: The clients are sampled / selected based on pk.
([pg. 4] Each client can compute the score sk using its received global model and send sk back to the server. On the server side, we then compute a probability pk for each client-k to be selected for aggregation [see below])
PNG
media_image7.png
95
706
media_image7.png
Greyscale
NOTE: Wang teaches the aforementioned prediction confidence values sk for client k are used to compute pk, which is used to determine which clients to select. Thus, Wang teaches selecting clients from the set of clients (sampling clients S from K) based on the prediction confidence values included in the updates (based on sk).
generating a new global model using the model updates in the updates received from the selected clients;
[pg. 4]
PNG
media_image8.png
63
1046
media_image8.png
Greyscale
and sending the new global model to each of the clients in the set of clients.
[pg. 4]
PNG
media_image9.png
17
456
media_image9.png
Greyscale
Regarding claim 8, Wang teaches;
the clients selected from the current training round may differ from clients selected during a different training round.
PNG
media_image10.png
112
570
media_image10.png
Greyscale
NOTE: The clients are selected according to the pk values, which are recalculated each training round. Thus, the clients selected from the current training round may differ from clients selected during a different training round, according to the newly calculated pk values.
Regarding claims 9 and 16,
Claims 9 and 16 are non-transitory storage medium claims directly corresponding to claims 1 and 8, respectively, and are rejected using the same reasoning.
Claim Rejections - 35 USC § 103
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) 2-3, 10-11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022) as applied to claim 1 above, and further in view of Zhiwei Ling et al. (hereinafter Ling) (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022).
Regarding claim 2, Wang fails to teach but Ling teaches;
wherein the prediction confidence values are average predicted confidence values generated by each of the clients over the current training round.
([pg. 3] one training round of FedEntropy involves 4 steps... once the selected devices finish their local training, in step 2 they only forward their soft labels to the server ... For a device with an index of k, assuming that it has lk samples, we calculate the data distribution of corresponding device by averaging all its soft labels)
NOTE: Ling teaches clients (devices) generating and forwarding their soft labels over a current training round, where soft labels for each data sample which indicate the predicted confidence or probability of each possible class for a sample, and averaging the soft labels generated by each of the clients over the current training round.
Thus, Ling teaches average predicted confidence values (averaged soft labels of local samples) generated by each of the clients over the current training round (the soft labels are generated by each of the clients over the current training round).
OBVIOSNESS TO COMBINE LING WITH WANG:
Ling and Wang are analogous art to the present disclosure as they both pertain to federated learning.
Wang computes a prediction confidence value for each labeled sample of each client k, and calculates a client level confidence value sk by summing the confidence values for each sample of client k, where the client level confidence value sk is used for client selection.
Similarly, Ling obtains prediction confidence values (soft label vectors) for each sample of each device / client, but averages the the confidence values (soft label vectors) of each sample of the corresponding client.
Ling additionally utilizes the averaged soft labels of the local model of each device to filter devices that are not beneficial for aggregation, allowing for various improvements in the Federated learning framework;
([Ling, pg. 2] in one training round of FedEntropy, the cloud firstly collects the soft labels from selected devices and judges their contributions to the overall entropy of soft labels... In this way, FedEntropy effectively and safely filters both useless and harmful devices for aggregation, thus leading to higher classification performance with less communication overhead.)
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to generate the client level confidence scores used for client selection in Wang by averaging prediction-confidence values generated from the pre-argmax softmax-layer outputs (soft labels) of current round local models, as taught by Ling, to determine and select clients that are beneficial for aggregation, thereby improving aggregation quality, classification performance, and resource usage.
Regarding claim 3, Wang fails to teach but Ling teaches;
wherein the average confidence values are determined from outputs of a pre-argmax layer of the local model generated during the training round.
([pg. 3] one training round of FedEntropy involves 4 steps... once the selected devices finish their local training, in step 2 they only forward their soft labels to the server ... Based on the outputs of the softmax layer of a local model, we collect all the soft labels of samples … For a device with an index of k, assuming that it has lk samples, we calculate the data distribution of corresponding device by averaging all its soft labels)
NOTE: The aforementioned average confidence values (average soft labels over all client samples) are determined from outputs of the softmax layer of the local model, where a softmax layer is a pre-argmax layer because its class-probability outputs are available before an argmax operation selects the class having the greatest probability.
Thus, Ling teaches that the average confidence values (averaged soft labels) are determined from outputs of a pre-argmax layer of the local model (the soft labels are generated from the softmax layer of the local model) generated during the training round (the soft labels are generated during the current FedEntropy training round).
OBVIOUSNESS: Using the same reasoning from claim 2, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to generate the client level confidence scores used for client selection in Wang by averaging prediction-confidence values generated from the pre-argmax softmax-layer outputs of current round local models, as taught by Ling, to select clients that are not harmful for aggregation, thereby improving aggregation quality, classification, performance, and lower communication overhead.
Regarding claims 10-11,
Claims 10-11 are non-transitory storage medium claims directly corresponding to claims 2-3, respectively, and are rejected using the same reasoning.
Claim(s) 4, 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022), further in view of Ling (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022) as applied to claim 3 above, and further in view of Leandro G. Barajas et al. (hereinafter Barajas) (US 20070288414 A1, 2007-12-13)
Regarding claim 4, Wang teaches;
Predicted confidence values, for each client
(same reasoning from claim 1)
Wang fails to teach but Ling teaches;
Average predicted confidence values, for each client
(same reasoning from claim 2)
Wang and Ling fail to teach but Barajas teaches;
wherein each of the
([0025] Each prediction algorithm may be assigned a probability of correctly predicting... Each prediction algorithm can also be assigned a confidence interval associated with the prediction and its probability.)
NOTE: Barajas teaches a predicted confidence value (probability of correctly predicting) associated with a confidence interval for each prediction algorithm of a plurality of prediction algorithms.
OBVIOUSNESS TO COMBINE BARAJAS WITH WANG AND LING:
Barajas is analogous art to the present disclosure as it pertains to assigning a predicted confidence value and associated confidence interval to a plurality of entities.
Wang teaches a federated learning framework in which client level predicted confidence values are computed and used to select clients,
Ling teaches a federated learning framework in which confidence values for each client are averaged, and are then used to filter clients that are not useful for aggregation,
And Barajas teaches associating a confidence interval with predicted confidence values of prediction algorithms.
Confidence intervals for predicted confidence values indicate the uncertainty of the predicted confidence values.
Accordingly, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to provide a confidence interval for each client level average predicted confidence value of Wang as modified by Ling, predictably enabling the server to account for the uncertainty of the predicted average confidence values, thereby improving the reliability of client selection and reducing the influence of clients whose apparently favorable predicted average confidence scores are less statistically reliable.
Regarding claim 12,
Claim 12 is a non-transitory storage medium claim directly corresponding to claim 4, and is rejected using the same reasoning.
Claim(s) 5, 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022), further in view of Ling (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022), further in view of Barajas (US 20070288414 A1, 2007-12-13) as applied to claim 4 above, and further in view of Titao Chen et al. (hereinafter Chen) (“CONFIDENCE-BASED FEDERATED DISTILLATION FOR VISION-BASED LANE-CENTERING,” 2023)
Regarding claim 5, Wang teaches;
prediction confidence values
same reasoning from claim 1
Wang fails to teach but Ling teaches;
average prediction confidence values
same reasoning from claim 2
Wang and Ling fail to teach but Barajas teaches;
confidence intervals associated with prediction confidence values
same reasoning from claim 4
Wang, Ling, and Barajas fail to teach but Chen teaches;
sorting the clients based on the
([Abstract] determine the predictive confidence of each local model)
[pg. 3]
PNG
media_image11.png
428
516
media_image11.png
Greyscale
NOTE: Chen teaches each client training and returning one local / teacher model, then sorting the teacher models associated with each client based on prediction confidence values (thus, sorting the clients based on the confidence score associated with their local model).
OBVIOUSNESS TO COMBINE CHEN WITH WANG, LING, AND BARAJAS:
Chen is analogous art to the present disclosure as it pertains to federated learning using confidence values.
Wang teaches a federated learning framework in which client level predicted confidence values are computed and used to select clients,
Ling teaches a federated learning framework in which confidence values for each client are averaged, and are then used to filter clients that are not useful for aggregation,
Barajas teaches associating a confidence interval with predicted confidence values of prediction algorithms,
And Chen teaches sorting clients based on prediction confidence values for corresponding local models, to select higher quality models to update the global / server model.
Thus, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to sort the clients of Wang as modified by Ling and Barajas based on the prediction confidence values of the local models corresponding to the clients as taught by Chen to provide a more direct and reliable mechanism for identifying the most confident clients to be selected, thereby reducing the likelihood for selecting poor local models for aggregation, and predictably improving the robustness and performance of the global model.
Regarding claim 13,
Claim 13 is a non-transitory storage medium claim directly corresponding to claim 5, and is rejected using the same reasoning.
Claim(s) 6-7, 14-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022) in view of Ling (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022) as applied to claim 1 above, and further in view of Ferdinand Kahenga et al. (hereinafter Kahenga) (“FedFaSt: Selective Federated Learning using Fittest Parameters Aggregation and Slotted Clients Training,” 2023).
Regarding claim 6, Wang teaches;
prediction confidence values
same reasoning from claim 1
Wang fails to teach but Ling teaches;
average prediction confidence values
same reasoning from claim 2
Wang and Ling fail to teach but Kahenga teaches;
selecting n clients whose [fitness]
([pg. 2] FedFaSt uses a threshold-aware fitness assessment of clients. During the current round, the clients with performance above a threshold are elected as fit to enter the training and aggregation team for the next slot; and those with performance below the threshold are left aside for fitness re-evaluation at the end of the current slot.)
OBVIOUSNESS TO COMBINE KAHENGA WITH WANG AND LING:
Kahenga is analogous art to the present disclosure as it pertains to a federated learning architecture using a threshold value for client selection.
Wang teaches federated learning client selection based on client confidence values,
Ling teaches averaging confidence values for each client to determine whether they should used in aggregation,
and Kahenga teaches federated learning client selection based on a comparison of client performance values and a threshold.
Both the confidence values of Wang (and average confidence values of Ling) and the performance values of Kahenga indicate the reliability of the contribution of a given client. Applying the threshold comparison of Kahenga to the average confidence scores of Wang as modified by ling would provide a deterministic quality cutoff, excluding un-reliable clients from being selected.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to apply the threshold comparison of Kahenga to the average confidence scores of Wang as modified by Ling to provide a deterministic quality cutoff, thereby excluding un-reliable clients from being selected for global aggregation, predictably improving the robustness and convergence of the federated learning process.
Thus, Wang, ling, and Kahenga reasonably teach the following claim in combination;
selecting n clients [Wang] whose average prediction confidence [Ling] values are greater than a threshold value [Kahenga]
Regarding claim 7, Wang teaches
predicted confidence values of the selected clients
same reasoning from claim 1
Wang fails to teach but Ling teaches
average predicted confidence values
same reasoning from claim 2
Wang and Ling fail to teach but Kahenga teaches;
updating the threshold value after the training round for a next training round based on an average of the [scores]
[pg. 2]
PNG
media_image12.png
59
456
media_image12.png
Greyscale
NOTE: Kahenga teaches updating the threshold value after the training round t for a next training round t+1 based on an average of the scores of all clients, including the scores of the selected clients.
OBVIOUSNESS:
Using the same reasoning from claim 6, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to apply the threshold comparison of Kahenga to the average confidence scores of Wang as modified by Ling to provide a deterministic confidence cutoff, thereby excluding un-reliable clients from being selected for global aggregation, predictably improving the robustness and convergence of the federated learning process.
Additionally, Wang recalculates the prediction confidence values for the clients each round, Ling teaches average confidence score values, and Kahenga defines their client fitness threshold as a function of the client scores at the current training round.
Because the prediction confidences of Wang and Ling change as federated training progresses, a person of ordinary skill would have understood that a fixed confidence threshold could become stale and cease to accurately distinguish reliable clients from unreliable clients.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to recalculate the threshold after each training round, as taught by Kahenga, based on the current average confidence scores of the clients taught by Wang as modified by Ling, to calibrate the threshold to the most recent average prediction confidence values, predictably maintaining an accurate confidence cutoff for selecting clients.
Regarding claims 14-15,
Claims 14-15 are non-transitory storage medium claims directly corresponding to claim 6-7, respectively, and are rejected using the same reasoning.
Claim(s) 17, 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022), further in view of Ling (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022), further in view of Barajas (US 20070288414 A1, 2007-12-13) further in view of Kahenga (“FedFaSt: Selective Federated Learning using Fittest Parameters Aggregation and Slotted Clients Training,” 2023).
Regarding claim 17, Wang teaches;
A method comprising: performing federated learning in a federated learning system that includes a set of clients and a central server, wherein each training round includes a first phase [client confidence value determination], a second phase [client selection based on confidence values], and a third phase [aggregating local model updates and generating a new global model]
([pg. 4] Given the confidence scores for all samples in Dk, we can estimate the number of clean labels on client-k as sk by summing up the confidence scores of all the samples on the client. Each client can compute the score sk using its received global model and send sk back to the server. On the server side, we then compute a probability pk for each client-k to be selected for aggregation)
NOTE: First phase; client confidence value sk determination.
[pg. 4]
PNG
media_image13.png
247
547
media_image13.png
Greyscale
NOTE: Second phase, client selection based on pk values, which are determined using the aforementioned confidence values sk.
[pg. 3]
PNG
media_image4.png
125
575
media_image4.png
Greyscale
NOTE: Third phase; aggregating local model updates and generating a new global model.
performing the first phase at each of the clients in the set of clients, wherein the first phase includes determining an
([pg. 4] Each client can compute the score sk using its received global model and send sk back to the server.)
NOTE: Wang teaches determining the aforementioned prediction confidence values sk at each of the clients in the set of clients.
performing the second phase at the central server, wherein the second phase includes selecting n clients from the set of clients
[pg. 4]
PNG
media_image14.png
77
751
media_image14.png
Greyscale
NOTE: pk is derived using the aforementioned prediction confidence values sk.
[pg. 4]
PNG
media_image15.png
24
504
media_image15.png
Greyscale
NOTE: Wang teaches performing, at the central server, selection of n clients (the number of clients in S) from the set of clients based on the prediction confidence values (according to pk, which is derived from the aforementioned prediction confidence values, sk).
performing the third phase at the central server, wherein the third phase includes aggregating local model updates received from the n clients and generating a new global model; and distributing the new global model to all of the clients in the set of clients
[pg. 3]
PNG
media_image9.png
17
456
media_image9.png
Greyscale
and repeating the first phase, the second phase, and the third phase.
[pg. 4]
PNG
media_image13.png
247
547
media_image13.png
Greyscale
NOTE: Wang teaches the process (including the previously mentioned first, second, and third phase) being repeated for R iterations.
Wang fails to teach but Ling teaches;
average prediction confidence
(using the same reasoning from claim 2)
OBVIOUSNESS:
Using the same reasoning from claim 2, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to generate the client level confidence scores used for client selection in Wang by averaging prediction-confidence values generated from the pre-argmax softmax-layer outputs of current round local models, as taught by Ling, to select clients that are not harmful for aggregation, thereby improving aggregation quality, classification, performance, and lower communication overhead.
Wang and Ling fail to teach but Barajas teaches;
-confidence interval
(using the same reasoning from claim 4)
OBVIOUSNESS:
Using the same reasoning from claim 4, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to provide a confidence interval for each client level average predicted confidence value of Wang as modified by Ling, predictably enabling the server to account for the uncertainty of the predicted average confidence values, thereby improving the reliability of client selection and reducing the influence of clients whose apparently favorable predicted average confidence scores are less statistically reliable.
From this, the combination of Wang, Ling, and Barajas reasonably teach;
performing the first phase at each of the clients in the set of clients [Wang], wherein the first phase includes determining an average prediction confidence value [Ling] and a confidence interval [Barajas]
Wang, Ling, and Barajas fail to teach but Kahenga teaches;
selecting n clients from the set of clients whose [fitness]
(using the same reasoning from claim 6)
OBVIOUSNESS:
Using the same reasoning from claim 6, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to apply the threshold comparison of Kahenga to the average confidence scores of Wang as modified by Ling to provide a deterministic quality cutoff, thereby excluding un-reliable clients from being selected for global aggregation, predictably improving the robustness and convergence of the federated learning process.
From this, in combination, Wang, Ling, and Kahenga reasonably teach;
selecting n clients from the set of clients [Wang] whose average prediction confidence [Ling] values are greater than a threshold value [Kahenga]
Regarding claim 19, Wang fails to teach but Ling teaches;
the average prediction confidence values are generated from an output of a pre-argmax layer of the model.
Using the same reasoning from claim 3
OBVIOUSNESS:
Using the same reasoning from claim 2, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to generate the client level confidence scores used for client selection in Wang by averaging prediction-confidence values generated from the pre-argmax softmax-layer outputs of current round local models, as taught by Ling, to select clients that are not harmful for aggregation, thereby improving aggregation quality, classification, performance, and lower communication overhead.
Claim(s) 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022), further in view of Ling (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022), further in view of Barajas (US 20070288414 A1, 2007-12-13) further in view of Kahenga (“FedFaSt: Selective Federated Learning using Fittest Parameters Aggregation and Slotted Clients Training,” 2023) as applied to claim 17 above, further in view of Nurettin Burcak Beser et al. (hereinafter Beser) (US 20190012592 A1, 2019-01-10).
Regarding claim 20, Wang teaches
prediction confidence values transmitted to the central server
([pg. 4] Given the confidence scores for all samples in Dk, we can estimate the number of clean labels on client-k as sk by summing up the confidence scores of all the samples on the client. Each client can compute the score sk using its received global model and send sk back to the server.)
Wang fails to teach but Ling teaches;
average prediction confidence values
using the same reasoning from claim 2
Wang, Ling, and Barajas fail to teach but Kahenga teaches;
and wherein the local model updates comprise model weights.
([pg. 1] Federated Learning (FL) has emerged as a groundbreaking paradigm for preserving privacy in collaborative learning. In this approach, individual clients (devices) create local models using their respective data and communicate weight updates to a central server.)
OBVIOUSNESS:
Kahenga directly indicates that in federated learning, local model updates comprise model weights. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, for the local model updates of the federated learning framework of Wang to comprise model weights.
Wang, Ling, Barajas, and Kahenga fail to teach but Beser teaches;
protecting privacy of the clients at least by encrypting [data]
([0025] When messages are sent between the central server and the ANNs, they are, in various embodiments, authenticated and encrypted using cryptographic protocols.)
OBVIOUSNESS TO COMBINE BESER WITH WANG, LING, BARAJAS, AND KAHENGA: Beser is analogous art to the present disclosure as it pertains to secure federated learning.
Wang teaches transmitting confidence values from local models to a central server in a federated learning framework, Ling teaches averaging confidence values for clients in a federated learning framework, and Beser teaches encrypting data sent between local models and a central server in a federated learning framework.
Additionally, Beser indicates that encrypting data sent between the local models and the server ensures only the intended recipients can decipher the communications, predictably protecting privacy of the clients;
([Beser, 0025] When messages are sent between the central server and the ANNs, they are, in various embodiments, authenticated and encrypted using cryptographic protocols… encryption ensures only the central server and particular ANNs can decipher that communication.)
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, for the average prediction confidence values of Wang as modified by Ling to be encrypted prior to transmission to the central server (as taught by Beser) to protect the privacy of the clients.
From this, the combination of Wang, Ling, and Beser reasonably teaches;
protecting privacy of the clients at least by encrypting [Beser] at least the average prediction confidence values [Ling] prior to transmission to the central server [Wang]
Claim(s) 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang (“FedNoiL: A Simple Two-Level Sampling Method for Federated Learning with Noisy Labels,” 20 May 2022), further in view of Ling (“FedEntropy: Efficient Device Grouping for Federated Learning Using Maximum Entropy Judgment,” 24 May 2022), further in view of Barajas (US 20070288414 A1, 2007-12-13) further in view of Kahenga (“FedFaSt: Selective Federated Learning using Fittest Parameters Aggregation and Slotted Clients Training,” 2023) as applied to claim 17 above, further in view of Vasileios Tsouvalas et al. (hereinafter Tsouvalas) (“Federated Self-Training for Semi-Supervised Audio Recognition,” 2022-02-25).
Regarding claim 18, Wang teaches;
wherein the training rounds are repeated until a loss converges or other stopping criteria is satisfied.
[pg. 4]
PNG
media_image16.png
146
557
media_image16.png
Greyscale
NOTE: Wang teaches a federated learning algorithm in which training rounds r are repeated until a stopping criteria is satisfied (the number of rounds, R, is reached).
Wang, Ling, Barajas, and Kahenga fail to teach but Tsouvalas teaches;
increasing the threshold value after each training round,
([pg. 12-13] we set the confidence threshold 𝜏 to initialize from 0.5 and gradually increase to a maximum of 0.9 during training, following a cosine schedule.)
OBVIOUSNESS TO COMBINE TSOUVALAS:
Tsouvalas is analogous art to the present disclosure as it pertains to confidence value-based thresholding in a federated learning framework.
Wang teaches a stopping condition for training rounds, as well as computing prediction confidence values to select clients for more robust global aggregation, Kahenga teaches comparing client fitness scores to a dynamically changing threshold value for client selection, and Tsouvalas teaches an increasing prediction confidence-based threshold value for federated learning.
Tsouvalas additionally indicates that early in training, when models are immature, applying a threshold value that is too high could exclude useful data, and during later training rounds, after predictions have become more reliable, maintaining a low threshold could cause training to utilize less reliable information. Increasing the threshold value progressively during training prevents these issues.
([Tsouvalas, pg. 10] with our cosine confidence thresholding, we allow clients to explore the locally-stored unlabeled data, 𝐷𝑘 𝑈, in the first few federated rounds, while considering only highly-confident predictions in a later stage of the training procedure.)
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the confidence based client selection threshold value taught by Wang as modified by Kahenga by increasing the threshold value over the training rounds (as taught by Tsouvalas), allowing for more exploration with client selection in early rounds and stricter confidence requirements as training continues, with the predictable result of progressively increasing the reliability of the information admitted into global aggregation.
CONCLUSION
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Matthew Alan Cady whose telephone number is (571) 272-7229. The examiner can normally be reached Monday - Friday, 7:30 am - 5:00 pm ET.
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, Cesar Paula can be reached on (571)272-4128. 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.
/MATTHEW ALAN CADY/ Examiner, Art Unit 2145
/CESAR B PAULA/ Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights