DETAILED ACTION
This Office Action is sent in response to the Applicant’s Communication received on 04/29/2026 for application number 17/924,991. The Office hereby acknowledges receipt of the following and placed of record in file: Specification, Drawings, Abstract, Oath/Declaration, IDS, and Claims.
Claims 2 and 5-8 are canceled.
Claims 1 is currently amended.
Claims 1, 3, 4, 9-11 are pending.
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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 04/29/2026 has been entered.
Response to Arguments
35 USC 101
On page 9 of the remarks section, Applicant argues that Based on the features of "applying the generated time series data adversarial sample with global disturbance to various time series data prediction models to perform adversarial attacks to reduce the prediction accuracy of the models" in claim 1, when the time series data prediction models adopts the generated time series data adversarial sample with global disturbance, the performance of the time series data prediction models can be effectively reduced, so that the reasoning of information in time series data by the time series data prediction models can be suppressed. That is, an effective adversarial attack solution is provided for the time series data prediction in the industrial field, and the solution has important significance for safe application of an industrial system and has wide applicability and transferability. The solution in claim 1 includes specific technical structures and data processing steps, and produces actual technical improvements that are rooted in the technical features, not merely in the environment or in mental steps. These technical elements are not generic computer functions, routine operations, or mere field-of-use limitations. They provide a non-conventional technical solution and constitute significantly more than any underlying abstract concept.
Upon further review, in view of the newly amended claimed limitation, the Examiner finds the Applicant’s argument persuasive. Therefore, the 35 USC 101 rejection is withdrawn
35 USC 103
On page 6 of the remarks section, the Applicant argues that Xiao only teaches calculating the gradient of each parameter based on the loss result of the proxy model and the weight of the parameter itself (which is different from the distance between the first importance and the second importance), and then the gradient of each parameter is used as the importance score of the corresponding parameter.
Applicant’s arguments have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of reference He.
The Applicant further argues that it is claimed that the distance between the first importance at the time instant and the second importance at the time instant is quantified by using Frobenius norm, which is a non-obvious technical choice that confers significant, unexpected technical advantages-advantages that are not suggested by Xiao or any other Cited Art.
Applicant’s arguments related to the cited limitation have been considered but are moot because the newly amended claim necessitated a new ground of rejection that does not rely on the previously cited prior art alone. Rather, Wei-Xiong-Hu-Chen-Xiao in combination with newly applied prior art, Liu(2), was used to address the said limitation.
On page 7 of the remarks section, Applicant argues that, In Xiao, only the importance of the parameters of the proxy model on the generated adversarial examples is considered. Then, some parameters of the proxy model are pruned based on preset rules and the importance scores of each parameter of the proxy model, and the final adversarial sample is generated. Xiao does not consider the importance in the time series data adversarial sample (the first importance) and the importance in the original time series data (the second importance). In contrast, the claimed invention's method is a multi-step, novel process that leverages the difference between the adversarial sample's importance and the original data's importance to generate a local disturbance sample-a type of adversarial sample that Xiao does not disclose or suggest. This local disturbance approach addresses a key limitation of global disturbance samples (e.g., over-disturbance that reduces real-world applicability) by targeting only the most impactful time instants, resulting in more effective, realistic adversarial samples that better test model robustness. Xiao's parameter-pruning approach does not achieve this technical effect, as it focuses on model parameters rather than time series data itself.
The Examiner respectfully disagrees. The claimed language is broad. Therefore, under the Broadest Reasonable Interpretation (BRI) of the claim language, Xiao does indeed consider the importance in the time series data adversarial sample (the first importance) and the importance in the original time series data (the second importance). Specifically, in paragraph 0015, Xiao teaches: “Calculate (calculating) the gradient of each parameter based on the loss result of the proxy model and the weight of the parameter itself”. In paragraph 0016, Xiao also teaches: “The gradient of each parameter (at each of time instants) is used as the importance score of the corresponding parameter (a first importance… in the adversarial sample). The importance score is used to indicate the degree of influence of the corresponding parameter of the proxy model (a second importance at each of time instants in the original data) on the generated adversarial sample in the corresponding round”. Additionally, the alleged “generat[ing] a local disturbance sample”, “local disturbance approach addresses a key limitation of global disturbance samples”, and “over-disturbance that reduces real-world applicability by targeting only the most impactful time instants, resulting in more effective, realistic adversarial samples that better test model robustness” are not part of the claim language. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
On page 8 of the remarks section, Applicant further argues that the cited Art (particularly Xiao) does not disclose the core Distinct Features- including the two separate importance values, the Frobenius norm-based distance calculation, and the local disturbance generation process. Xiao's focus on proxy model parameters is unrelated to the claimed invention's focus on time instant importance in adversarial vs. original time series data; there is no suggestion in Xiao (or any other Cited Art) that comparing importance values between two datasets (adversarial and original) would be useful, let alone that Frobenius norm should be used to quantify that comparison. In determining the final adversarial sample, the solution for determining the importance of the parameters and generating the time series data adversarial sample according to claim 1 is completely different from the solution disclosed by Xiao, and the resulting improvements to the model output are also different. Xiao fails disclosing or teaching the technical features of claim 1. Those skilled in the art cannot derive the above technical features based on BRI, Xiao, and the other art documents.
The Examiner respectfully disagrees. As mentioned above, first, under the Broadest Reasonable Interpretation (BRI) of the claim language, Xiao does indeed consider the importance in the time series data adversarial sample (the first importance) and the importance in the original time series data (the second importance). Specifically, in paragraph 0015, Xiao teaches: “Calculate (calculating) the gradient of each parameter based on the loss result of the proxy model and the weight of the parameter itself”. In paragraph 0016, Xiao also teaches: “The gradient of each parameter (at each of time instants) is used as the importance score of the corresponding parameter (a first importance… in the adversarial sample). The importance score is used to indicate the degree of influence of the corresponding parameter of the proxy model (a second importance at each of time instants in the original data) on the generated adversarial sample in the corresponding round”. Second, Applicant’s arguments related to the Frobenius norm-based distance calculation have been considered but are moot because the newly amended claim necessitated a new ground of rejection that does not rely on the previously cited prior art alone. Rather, Wei-Xiong-Hu-Chen-Xiao in combination with newly applied prior art, Liu(2), was used to address the said limitation. Third, the alleged “generat[ing] a local disturbance sample”, “local disturbance approach addresses a key limitation of global disturbance samples”, and “over-disturbance that reduces real-world applicability by targeting only the most impactful time instants, resulting in more effective, realistic adversarial samples that better test model robustness” are not part of the claim language. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993). Furthermore, in response to Applicant’s argument that “Xiao's focus on proxy model parameters is unrelated to the claimed invention's focus on time instant importance in adversarial vs. original time series data”, Applicant appears to argue cited Xiao’s teaching not meeting the claimed invention based on Applicant’s disclosed invention. Moreover, there is no explicit claim limitation related to “useful”. In response to claim 1 being completely different from the solution disclosed by Xiao, Examiner respectfully states that the said claim language is broad, and under the BRI, Xiao indeed meets the said limitations.
Finally, Applicant argues that, based on the above amended claim 1, the time series data adversarial sample with local disturbance is generated based on the distance parameter, so that the output results of the conventional model are improved, which achieves technical effects and improving the conventional model.
The Examiner respectfully states that the analyses considering technical effects and improvements is part of the 35 USC 101 analysis under abstract idea, not 35 USC 103 prior art consideration.
Therefore the 35 USC 103 rejection is maintained using a new ground of rejection.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1, 9, and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Wei et al. (WO 2020046388 A1), hereinafter Wei, in view of Xiong et al. (CN112257851A, see attached translation), hereinafter Xiong, Hu et al. (CN112232495A, see attached translation), hereinafter Hu, Chen et al. (US 20210067549 A1), hereinafter Chen, Xiao et al. (CN112329930A, see attached translation), hereinafter Xiao, Liu et al (CN110443117A, see attached translation), hereinafter Liu(2), and He et al. (CN112418325A, see attached translation), hereinafter He.
Regarding claim 1, Wei teaches,
A method for generating a time series data adversarial sample [Abstract, Systems, techniques, and computer-program products are provided to generate synthetic time series using a generative adversarial network], comprising: superimposing the noise (Para 19, random noise data 104) on the original time series data (Para 17, observed time series) to generate (Para 19, create) a time series data adversarial sample with global disturbance (Para 19, realistic time series) [Para 17, embodiments of the disclosure apply adversarial training to observed time series (or time series data) in order to configure the GAN to generate a synthetic time series; Para 19, the generator network module can create a synthetic time series 114 from random noise data 104 in an attempt to generate a realistic time series].
Wei does not teach Training a time series prediction model by using original time series data; calculating a maximum value of a loss function in the time series prediction model based on a stochastic gradient descent optimization algorithm; and determining a noise based on the maximum value of the loss function; and applying the generated time series data adversarial sample with global disturbance to various time series data prediction models to perform adversarial attacks to reduce the prediction accuracy of the models; Calculating a first importance at each of time instants in the adversarial sample and a second importance at each of time instants in the original data; calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant, wherein the distance between the first importance at the time instant and the second importance at the time instant is quantified by using Frobenius norm; sorting distances at all corresponding time instants in a descending order to determine first several time instants; and replacing data at corresponding time instants in the original data to generate an adversarial sample with local disturbance..
Xiong teaches,
Calculating a value of a loss function (Para 0041, adding a gradient penalty term of the loss) based on a stochastic gradient descent optimization algorithm (Para 0041, stochastic gradient descent (Sgd) algorithm); determining a noise (Para 0041, adds adversarial perturbations) based on the value of the loss function; [Para 0041, The gradient descent algorithm can be a stochastic gradient descent (Sgd) algorithm… This embodiment adds adversarial perturbations to the input samples for equivalent conversion by applying the gradient descent method, and achieves the purpose of adversarial training by adding a gradient penalty term of the loss with respect to the input to the loss function.]
Xiong is analogous to the claimed invention as they both relate to adversarial training. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Xiong and provide a loss function based on SGD for its computational efficiency when calculating gradients, and to provide determination of a noise in order to perturb data for adversarial training.
Wei-Xiong do not teach Training a time series prediction model by using original time series data; wherein the value of a loss function is a maximum value of a loss function; applying the generated time series data adversarial sample with global disturbance to various time series data prediction models to perform adversarial attacks to reduce the prediction accuracy of the models; Calculating a first importance at each of time instants in the adversarial sample and a second importance at each of time instants in the original data; calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant, wherein the distance between the first importance at the time instant and the second importance at the time instant is quantified by using Frobenius norm; sorting distances at all corresponding time instants in a descending order to determine first several time instants; and replacing data at corresponding time instants in the original data to generate an adversarial sample with local disturbance..
Hu teaches,
Training (Para 0016, training) a time series (Para 0015, data includes time series data) prediction model (Para 0016, prediction model) by using original time series data (Para 0015, the original data includes time series data) [Para 0015-0017, Get the original data… the original data includes time series data, the preset method further includes segmenting the time series data according to a time window format… training the prediction model based on the acquired training data with the goal of converging the network weights of the prediction model includes: Inputting the acquired training data into the distribution estimator of the prediction model;];
wherein the value of a loss function is a maximum value (Para 0024, maximum mean difference) of a loss function [Para 0024, The distribution matcher in the prediction model is controlled to calculate the data sample through the maximum mean difference loss function so that the network weight of the prediction model converges again].
Hu is analogous to the claimed invention as they both relate to utilizing time series prediction models. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei and Xiong’s teachings to incorporate the teachings of Hu and provide a training a time series prediction model utilizing original data in order to provide meaningful data to the adversarial training model. It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to provide the value of a loss function being a maximum value in order to train input samples that have the most significant effect on a model.
Wei-Xiong-Hu teach the above limitations of claim 1 including the time series data adversarial sample with global disturbance (Wei, para 19).
Wei-Xiong-Hu do not teach applying adversarial sample to various time series data prediction models to perform adversarial attacks to reduce the prediction accuracy of the models; Calculating a first importance at each of time instants in the adversarial sample and a second importance at each of time instants in the original data; calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant, wherein the distance between the first importance at the time instant and the second importance at the time instant is quantified by using Frobenius norm; sorting distances at all corresponding time instants in a descending order to determine first several time instants; and replacing data at corresponding time instants in the original data to generate an adversarial sample with local disturbance.
Chen teaches,
applying adversarial sample to various time series data prediction models to perform adversarial attacks to reduce the prediction accuracy of the models
[Para 0028, The parameters of a GNN may be trained in a supervised manner, with the aim being to learn a new representation of a node, which can be used to predict the node label; Para 0029, During an adversarial attack, the graph is maliciously transformed through a perturbation function to create a perturbed graph G′, such that the performance of the trained GNN model drops significantly. To prepare the model for such attacks, the present embodiments refine the graph in block 306 using adversarial samples that simulate an adversarial attack; Para 0076, Referring now to FIG. 9, the overall structure of the GNN is shown. The refined graph encoder 902 accepts graph samples (including original samples and perturbed adversarial samples) as an input. The samples are processed by a series of K aggregation/updating steps, with a first aggregator 904 aggregating the neighbors of an unlabeled node and first updater 906 determines updated representations of each node, producing a representation h.sub.v.sup.1 908… A final, K.sup.th stage includes a K.sup.th aggregator 910 and a K.sup.th updater 914, producing a K.sup.th representation 916. Each aggregator may be implemented as a mean aggregator, and each updater may be implemented as a densely connected updater].
Chen is analogous to the claimed invention as they both relate to generating adversarial samples. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Chen and provide applying an adversarial samples to multiple prediction models in order to bolster the robustness of the model’s accuracy.
Wei-Xiong-Hu-Chen teach the limitations of claim 1 including wherein after generating the time series data adversarial sample with global disturbance, the method further comprises, time series data, and the generated adversarial sample with global disturbance.
Wei-Xiong-Hu-Chen do not teach Calculating a first importance at each of time instants in the adversarial sample and a second importance at each of time instants in the original data; calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant, wherein the distance between the first importance at the time instant and the second importance at the time instant is quantified by using Frobenius norm; sorting distances at all corresponding time instants in a descending order to determine first several time instants; and replacing data at corresponding time instants in the original data to generate an adversarial sample with local disturbance.
Xiao teaches,
Calculating (Para 0015, Calculate) a first importance at each of time instants (Para 0016, each parameter) in the adversarial sample (Para 0016, importance score… on the generated adversarial sample) and a second importance (Para 0016, the degree of influence of the corresponding parameter of the proxy model) at each of time instants in the original data (Para 0016, proxy model) [Para 0015, Calculate the gradient of each parameter based on the loss result of the proxy model and the weight of the parameter itself; Para 0016, The gradient of each parameter is used as the importance score of the corresponding parameter. The importance score is used to indicate the degree of influence of the corresponding parameter of the proxy model on the generated adversarial sample in the corresponding round.];
calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant [Para 0015, Calculate the gradient of each parameter based on the loss result of the proxy model and the weight of the parameter itself; Para 0016, The gradient of each parameter is used as the importance score of the corresponding parameter. The importance score is used to indicate the degree of influence of the corresponding parameter of the proxy model on the generated adversarial sample in the corresponding round.];
sorting (Para 0115, sorted) distances (Para 0115, importance score) at all corresponding time instants (Para 0115, of each parameter) in a descending order (Para 0115, in descending order) to determine first time instants (Para 0115, remaining parameters after pruning 20%) [Para 0115, based on the importance score of each parameter, each parameter is sorted in descending order, and the parameters in the bottom 20% are pruned];
Xiao is analogous to the claimed invention as they both relate to adversarial learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Xiao and provide calculating and sorting importances [Xiao, Para 0061] in order to pay more attention to important features which improves output results of the model.
Wei-Xiong-Hu-Chen and the cited embodiment of Xiao teach the limitations of claim 1 and the above limitations including the first time instants (see above).
Wei-Xiong-Hu-Chen and the cited embodiment of Xiao do not teach, calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant, wherein the distance between the first importance at the time instant and the second importance at the time instant is quantified by using Frobenius norm; and replacing data at corresponding time instants in the original data to generate an adversarial sample with local disturbance.
Xiao in a different embodiment teaches,
replacing (Para 0123, inputs it into the proxy model) data at corresponding time instants (Para 0123, each parameter) in the original data (Para 0123, proxy model) to generate an adversarial sample with local disturbance (Para 0123, multiple loss results) [Para 0123, In another embodiment of the present embodiment, a different parameter gradient calculation scheme is proposed. In each iteration round, the scheme randomly samples according to a preset method to generate white noise that obeys a preset probability distribution, and superimposes the white noise on the adversarial sample generated in the round and inputs it into the proxy model to obtain corresponding multiple loss results. The gradient of each parameter is then calculated based on the multiple loss results to ensure that the importance score of each parameter is more stable.].
Xiao is analogous to the claimed invention as they both relate to adversarial learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Xiao and provide calculating and sorting importances [Xiao, Para 0061] in order to pay more attention to important features which improves output results of the model.
Wei-Xiong-Hu-Chen-Xiao teach the above limitations of claim 1 including the distance between the first importance at the time instant and the second importance at the time instant.
Wei-Xiong-Hu-Chen-Xiao do not teach calculating, at each of corresponding time instants, a distance between a first importance at the time instant and a second importance at the time instant, distance is quantified by using Frobenius norm.
Liu(2) teaches,
distance is quantified by using Frobenius norm [Pg. 7, para 1, the Frobenius norm is used as a matrix distance metric when aligning the covariance of source domain data and target domain data].
Liu(2) is analogous to the claimed invention as they both relate to time series data operations. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Liu(2) and provide using Frobenius norm in order to align and optimize the distance between datasets.
Wei-Xiong-Hu-Chen-Xiao- Liu(2) teach the above limitations of claim 1 including the time instant (Xiao, para 0015 and 0016).
Wei-Xiong-Hu-Chen-Xiao-Liu(2) do not teach calculating, at each of corresponding time instants, a distance between a first importance and a second importance.
He teaches,
calculating, at each of corresponding time instants (Abstract, each different variable), a distance between a first importance and a second importance (Abstract, This invention assigns a weight vector to each different variable, representing its importance) [Abstract, This invention proposes a soft subspace clustering method based on variable weighting, which solves the clustering problem of multivariate time series data and eliminates the influence of redundant or related variables among multivariate time series. This invention assigns a weight vector to each different variable, representing its importance to the formation of each cluster. To efficiently calculate the distance between two multivariate time series, a shape-based distance metric is first introduced].
He is analogous to the claimed invention as they both relate to processing time series data. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of He and provide a distance between importance values in order to create a metric that can iteratively help machine learning systems reach convergence.
Regarding claim 9, Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He teach the limitations of claim 1 including the method for generating a time series data adversarial sample.
Wei further teaches,
An electronic device, comprising: at least one processor (Para 44, one or more processors 350), and a memory coupled with the at least one processor (Para 45, The processor(s) 360 can be functionally coupled to the memory 370), wherein the memory stores a computer program (Para 46, machine-accessible instructions), and the computer program, when executed by the at least one processor (Para 46, machine-accessible instructions… that can be accessed and executed by at least one of the processor(s) 360), causes the processor to perform the method (Para 44, to generate synthetic time series in accordance with aspects of this disclosure) [Para 44, FIG. 3B presents an example of a computing system 350 to generate synthetic time series in accordance with aspects of this disclosure. As is illustrated in FIG. 3B, the computing system 350 can include one or more memory devices 370 (generically referred to as memory 370) that can retain or otherwise store the time series generation system 310. The computing system includes one or more processors 350; Para 45, The processor(s) 360 can be functionally coupled to the memory 370; Para 46, In the illustrated computing system 350, the time series generation system 310 can be embodied in or can include machine-accessible instructions (e.g., computer-readable and/or computer-executable instructions) that can be accessed and executed by at least one of the processor(s) 360].
Regarding claim 10, Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He teach the limitations of claim 1 including the method for generating a time series data adversarial sample.
Wei further teaches,
A non-transitory computer readable storage medium storing a computer program, wherein the computer program, when executed, performs the method [Para 65, to embody one such method, at least the portion of the computer- accessible instructions can be retained in a computer-readable storage non-transitory medium and executed by one or more processors].
Claim(s) 3 and 4 are rejected under 35 U.S.C. 103 as being unpatentable over Wei in view of Xiong, Hu, Chen, Xiao, Liu(2), and He, and in further view of Liu et al. (MI-FGSM on Faster R-CNN Object Detector, published 2020), hereinafter Liu, and Carmena et al. (US 20210365739 A1), hereinafter Carmena.
Regarding claim 3, Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He teach the limitations of claim 1 including the determining a noise based on the maximum value of the loss function.
Wei further teaches,
calculating a gradient of the loss function by using a symbolic function (Para 38, objective function V(Q, D )) [Para 38, performing a stochastic gradient descent (SGD) process… performance of the SGD process can implement a first-order gradient-based optimization of the objective function V(Q, D )];
Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He do not teach determining a linear noise parameter based on a maximum disturbance and an iteration number; and determining a maximum value of a product of the linear noise parameter multiplied by the calculated gradient as the noise.
Liu teaches,
determining (Algorithm 1, line 1) a linear noise parameter (Algorithm 1, line 1, a) based on a maximum disturbance (Algorithm 1, ϵ) and an iteration number (Algorithm 1, line 1, iterations T) [Sect 3, para 1, r is the perturbation, and ϵ is the allowed maximum value];
PNG
media_image1.png
415
586
media_image1.png
Greyscale
Liu is analogous to the claimed invention as they both relate to momentum iterative fast gradient sign method. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Liu and provide MIFGSM methodologies [Liu, Sect 1, pg. 28, col 1, para 2] in order to generate high quality adversarial examples by deriving from a strong attack method on classification.
Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He-Liu do not teach determining a maximum value of a product of the linear noise parameter multiplied by the calculated gradient as the noise.
Carmena teaches,
determining (Para 0063, to produce) a maximum value (Para 0063, level of magnitude of noise) of a product of the linear noise parameter (Para 0063, ϵ) multiplied (Para 0063, multiplied) by calculated gradient (Para 0063, calculated gradient) as the noise [Para 0063, a calculation containing elements corresponding to ϵ multiplied by elements corresponding to a calculated gradient sign can be used to produce a level of magnitude of noise].
Carmena is analogous to the claimed invention as they both relate to generating adversarial data. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei’s teachings to incorporate the teachings of Carmena and provide a maximum value of a product of the linear noise parameter multiplied by calculated gradient as the noise [Carmena, Para 0083] in order to offset data points in a plurality of categories, creating a variety of adversarial training data points which creates more robust ML models.
Regarding claim 4, Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He-Liu-Carmena teach the limitations of claim 3.
Liu further teaches,
wherein the linear noise parameter (Algorithm 1, line 1, a) is equal to a ratio of the maximum disturbance (Algorithm 1, line 1, ϵ) to the iteration number (Algorithm 1, line 1, T) [Sect 3, para 1, r is the perturbation, and ϵ is the allowed maximum value].
PNG
media_image2.png
426
601
media_image2.png
Greyscale
Claim(s) 11 is rejected under 35 U.S.C. 103 as being unpatentable over Wei, in view of Xiong, Hu, Chen, Xiao, Liu(2), and He, and in further view of Sun (CN109036389A, see attached translation), hereinafter Sun.
Regarding claim 11, Wei-Xiong-Hu-Chen-Xiao-Liu(2)-He teach the limitations of claim 1 including the calculating a maximum value of a loss function in the time series prediction model based on a stochastic gradient descent optimization algorithm (Xiong, para 0041; Hu, para 0024) and the maximum value of the loss function (Hu, para 0024).
Wei-Xiong-Hu-Chen-Xiao-Liu(2) do not teach determining, based on a direction opposite to a descent direction of a gradient, a direction along which the loss function increases fastest.
Sun teaches,
determining, based on a direction opposite to a descent direction of a gradient, a direction along which the loss function increases fastest [Para 0089, the input data of these samples will make the loss function increase in the fastest direction].
Sun is analogous to the claimed invention as they both relate to adversarial training. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wei, Xiong, and Hu’s teachings to incorporate the teachings of Sun and provide determining a direction along which the loss function increases fastest [Sun, para 0086] in order to quickly generate adversarial samples.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SYED RAYHAN AHMED whose telephone number is (571)270-0286. The examiner can normally be reached Mon-Fri 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, David Yi can be reached at (571) 270-7519. 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.
/SYED RAYHAN AHMED/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights