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 .
DETAILED ACTION
This action is responsive to the Application filed on 08/19/2026
Claims 1-11 and 14-17 are pending in the case. Claims 1 is the only independent claim. Claims 12-13 have been canceled.
Applicant’s arguments, see Response After Final action, filed on 08/19/2026, with respect to Final Rejection have been fully considered and are persuasive. The Non-Final rejection of 05/19/2026 has been withdrawn.
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.
Claims 1-2, 5-6, 8-9, 11 and 16-17 are rejected under 35 U.S.C 103 as being unpatentable over Mummadi et al. (Pub No.: 20180308012 A1), hereinafter referred to as Mummadi in view of “Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks”, https://arxiv.org/pdf/1704.01155, Xu et al, 12/05/2017, hereinafter referred to as Xu and further in view of Goswami et al. (Pub No.: 20190238568 A1), hereinafter referred to as Goswami.
With respect to claim 1, Xu disclose:
A method of operating a neural network-based sensing system for detecting and correcting for the presence of adversarial perturbations in the sensing system, the neural network-based sensing system including a processor coupled to a neural network module and operatively coupled to at least one sensor, the neural network module configured to implement a neural network, the method comprising at least the following steps: - receiving input signals (x) from the at least one sensor (In Fig. 1 and paragraph [0050], Mummadi discloses the receiving output signal S of the sensor in an optional receiving unit 50, which converts output signal S into a data signal x (alternatively, output signal S may also be directly adopted as data signal x).
Generating, from a plurality of the outputs including the given output yo, a measurement quantity (m), where m is, , (i) a first measurement quantity Mi as a value of a gradient Dxfϑ of the network function fo corresponding to the given input (xo), (ii) a second measurement quantity M2 corresponding to a gradient of a predetermined objective function derived from a training process for the neural network, or (iii) a third measurement quantity M3 derived from a combination of Mi and M2(Examiner selects (ii): (Under the broadest reasonable interpretation, the calculated gradient value constitutes the claimed "second measurement quantity" corresponding to a gradient of a predetermined objective function) In paragraph [0080-0083], Mummadi discloses generating a value corresponding to a gradient of a predetermined objective function derived from training of a neural network. Discusses a weighted cost function (e.g., cross entropy) for a machine learning system and further discloses ascertaining a gradient of the weighted cost function averaged over a training dataset.)
With respect to claim 1, Mummadi does not explicitly disclose:
Generating, from the input signals (x) including at least a given input (xo), respective outputs, the outputs being predictions of the neural network and including a given output yo corresponding to the given input (x0)
Detecting the presence of an adversarial perturbation in the sensing system in real time when the measurement quantity (m) is equal to or greater than a threshold
Executing, in response to detecting the presence of an adversarial perturbation, at least one remedial action in real time including: rejecting the output prediction of the neural network fϑ (x0) and stopping further action of the sensing system resulting from the rejected output prediction
Stopping the sensing system and issuing a warning via an output device
However, it is known by Xu to disclose:
Generating, from the input signals (x) including at least a given input (xo), respective outputs, the outputs being predictions of the neural network and including a given output yo corresponding to the given input (x0) (On page 2 (paragraph 6), Xu discloses generating, from respective input data including a given input x, respective outputs comprising predictions of a neural network. In particular, Xu discloses feeding input samples through a deep neural network and generating an output prediction g (xi) associated with the respective i-th input sample. Thus, Xu teaches a given neural network output corresponding to a given input.)
Detecting the presence of an adversarial perturbation in the sensing system in real time when the measurement quantity (m) is equal to or greater than a threshold (In Fig. 1 and page 1 (last paragraph), Xu teaches detecting the presence of an adversarial perturbation based on a measurement quantity satisfying a predetermined threshold. Specifically, in Fig.1, Xu teaches determining a measurement representing disagreement between the neural-network prediction for an original input and the prediction for a squeezed version of the input and the prediction for a squeezed version of the input and compares the resulting measurement against a selected threshold. When the disagreement exceeds the threshold, the input is identified and rejected as an adversarial example.)
Executing, in response to detecting the presence of an adversarial perturbation, at least one remedial action in real time including: rejecting the output prediction of the neural network fϑ (x0)
Mummadi and Xu are analogous pieces of art because both references concern method and device for adversarial examples. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Mummadi with improving robustness against adversarial examples with detecting adversarial examples, with detecting adversarial examples with high accuracy and few false positives as taught by Xu. The motivation for doing so would have been to reduce robustness against adversarial examples (See [0004] of Mummadi).
With respect to claim 1, Mummadi in view of Xu do not explicitly disclose:
Stopping further action of the sensing system resulting from the rejected output prediction
Stopping the sensing system and issuing a warning via an output device
However, it is known by Goswami disclose:
Stopping further action of the sensing system resulting from the rejected output prediction (In paragraph [0055], Goswami discloses, in response to detecting adversarial input data, a remedial or responsive operation including rejecting/discarding the adversarial input and preventing further action resulting therefrom. In paragraph [0112], Goswami discloses that once input is identified as adversarial, the input may be discarded and further processing thereof discontinued. Goswami further disclosed that in response to determining that input data is adversarial, the responsive operation may comprise shutting down the operation of a downstream computing system, thereby preventing downstream action based upon the detected adversarial input.))
Stopping the sensing system and issuing a warning via an output device (In paragraph [0093], Goswami discloses generating an alert by activating a siren and warning light or outputting a warning display on a computing terminal. In paragraph [0122], Goswami discloses in response to detecting adversarial input data, performing responsive remedial operation including shutting down operations of a downstream computing system and generating and outputting a notification indicating detection of the adversarial input.)
Mummadi in view of Xu and Goswami are analogous pieces of art because both references concern method and device for adversarial examples. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Goswami with identify one or both of classification values and categories having a high proportion of invalidated classifications as taught by Goswami. The motivation for doing so would have been to improve methods for incorporating both automated classification and human judgment in a highly effective manner. (See [0006] of Goswami.)
Regarding claim 2, Mummadi in view of Xu and Goswami disclose the elements of claim 1. In addition, Xu disclose:
The method according to claim 1, further comprising, if the measurement quantity (m) is determined to be less than the threshold, performing a predetermined usual action resulting from y.( On page 4 (2nd paragraph), Xu discloses that when the measured disagreement between neural network predictions does not satisfy the selected threshold for identifying an adversarial input, the input is considered legitimate, and the system outputs the neural network prediction for the legitimate input. Thus, Xu teaches, when the measurement quantity is below the adversarial detection threshold, proceeding with the predetermined usual operation associated with the neural network prediction, rather than rejecting the input as adversarial. )
Regarding claim 5, Mummadi in view of Xu and Goswami disclose the elements of claim 1. In addition, Xu disclose:
The method according to claim 1, wherein generating the second measurement quantity M2 comprises:- computing a gradient Da J(X, Y, fo) of the objective function by J(X, Y, fo) with respect to the network parameters B, whereby J(X, Y, fo) has been previously obtained by calibrating the network function fo in an offline training process based on given training data deriving the second measurement quantity M2 as the value of gradient Da J(X, Y, fo) corresponding to the given input (xo) (On page 3 (1) Fast Gradient Sign Method: GSM (L∞, Untargeted), Xu discloses training a DNA using labeled training data (xi, yi), wherein the difference between the network prediction g (xi) and the corresponding true label yi is modeled by an objective/loss function, and the loss id propagated backward through the network using back propagation to update the DNN parameter, thereby computing gradients of the training objective with respect to the network parameters)
Regarding claim 6, Mummadi in view of Xu and Goswami disclose the elements of claim 5. In addition, Xu disclose:
The method according to claim 5, wherein deriving the second measurement quantity M2 comprises determining the Euclidean norm of Da J(X, Y, fo) corresponding to the given input (xo) (On page 3 (second paragraph), Xu disclose The L2 norm corresponds to the Euclidean distance between x and x′.)
Regarding claim 8, Mummadi in view of Xu and Goswami disclose the elements of claim 1. In addition, Mummadi disclose:
The method according to claim 1, wherein the first measurement quantity Mi, the second measurement quantity M2 and/or the third measurement quantity M3 is generated based on a predetermined neighborhood of inputs (x) including the given input (xo) (In paragraphs[0052-0053], Mummadi discloses generating adversarial robustness -related quantities using regions, neighborhoods, surrounding pixels, and nearby input values associated with an image/input sample.)
Regarding claim 9, Mummadi in view of Xu and Goswami disclose the elements of claim 8. In addition, Mummadi disclose:
The method according to claim 1, wherein the first measurement quantity Mi, the second measurement quantity M2 and/or the third measurement quantity M3 is generated based on a predetermined neighborhood of inputs (x) including the given input (xo) (In paragraphs[0052-0053], Mummadi discloses generating adversarial robustness -related quantities using regions, neighborhoods, surrounding pixels, and nearby input values associated with an image/input sample.)
Regarding claim 11, Mummadi in view of Xu and Goswami disclose the elements of claim 1. In addition, Mummadi disclose:
The method according to claim 1, wherein the at least one remedial action includes saving the value of fo (xo) and wait for a next output fo(x1) in order to verify fa (xo) or to determine that it was a false output (In paragraph [003], Mermaid discloses comparing and verifying to determine whether the first output was false/adversarial.)
Regarding claim 14, Mummadi in view of Xu and Goswami disclose the elements of claim 2. In addition, Mummadi disclose:
A method of classifying outputs of a sensing system employing a neural network, the method comprising the method according to claim 2, wherein the predetermined usual action or the predetermined further actions comprise determining a classification or a regression based on the prediction y (In paragraph [0053], Mummadi discloses characterize that this pixel is classified as belonging to a semantic class (from a plurality of semantic classes). Semantic segmentation y_cls, may be a vectorial-valued variable for each pixel, which indicates, for each of the semantic classes with the aid of an associated number in the value interval [0;1], how high the probability is that this pixel is to be assigned to the respective semantic class.)
Regarding claim 16, Mummadi in view of Xu and Goswami disclose the elements of claim 1. In addition, Mummadi disclose:
A sensing and/or classifying system, for processing predictions and/or classifications in the presence of adversarial perturbations, the sensing and/or classifying system comprising:- a processor and, coupled thereto (In paragraph [0050], disclose a control system processing of output signal S.)
a memory (In paragraph [0088], Mermaid discloses storing data signals in memory)
wherein the processor is configured to connect to one or more sensors for receiving inputs (x) therefrom (In paragraph [0050], disclose receiving signal S)
wherein the processor is configured to run a module in the memory for implementing a neural network, the neural network having a network function fe, where B are network parameters (In paragraph [0059], Mermaid discloses neural-network functions and parameters.)
wherein the processor, is configured to execute the method of claim 1 (In paragraph [0004], Mermaid discloses adversarial perturbation generation and robustness analysis.)
Regarding claim 17, Mummadi in view of Kathrin and Jiajun Lu disclose the elements of claim 16. In addition, Mummadi disclose:
A vehicle comprising a sensing and/or classifying system according to claim 16 (In paragraph [0050], Mummadi discloses that the actuator control system 40 receives output signal S of the sensor.)
Claims 3-4 are rejected under 35 U.S.C 103 as being unpatentable over Mummadi in view of Xu, Goswami and further in view of “Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing their Input Gradients”, https://arxiv.org/pdf/1711.09404, Ross et al, 11/26/2017, hereinafter referred to as Ross.
Regarding claim 3, Mummadi in view of Xu and Goswami disclose the elements of claim 1. Mummadi in view of Xu and Goswami does not explicitly disclose:
The method according to claim 1, wherein generating the first measurement quantity Mi comprises:- computing the gradient Dxfo of the network function fo with respect to the input (x), and - deriving the first measurement quantity Mi as the value of gradient Dxfo corresponding to the given input (xo)
However, Ross disclose the limitation (On page 7 and Fig. 6, Ross discloses computing an input gradient of a differential neural network function with respect to input data. In particular, Ross teaches input-gradient regularization wherein gradients representing the rate of change of the neural network output/loss with respect to the input features are computed for respective input examples.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of Mummadi in view of Xu and Goswami before them, to include Ross, with a deep neural network less sensitive to small/adversarial changes in its inputs as taught by Ross. The motivation for doing so would have been to improve adversarial robustness and interpretability of DNN by regularizing their input gradients (See (Page 1) of Ross.)
Regarding claim 4, Mummadi in view of Xu, Goswami and Ross disclose the elements of claim 3. Mummadi in view of Xu, Goswami and Rossdoes not explicitly disclose:
The method according to claim 3, wherein deriving the first measurement quantity Mi comprises determining the Euclidean norm of Dxfo corresponding to the given input (xo) (On page 3 (second paragraph), Xu discloses the L2 norm corresponds to the Euclidean distance between x and x′.)
Claim 7 is rejected under 35 U.S.C 103 as being unpatentable over Mummadi in view of Xu, Goswami and further in view of Oberbreckling et al. (Pub No.: 20180074786 A1), hereinafter referred to as Oberbreckling.
Regarding claim 7, Mummadi in view of Xu and Goswami disclose the elements of claim 1. Mummadi in view of Xu and Goswami does not explicitly disclose:
The method according to claim 1, wherein the third measurement quantity M3 is computed as a weighted sum of the first measurement quantity Mi and the second measurement quantity M2
However, Oberbreckling disclose the limitation (Oberbreckling [0159], a third measure being the weighted sum of the first measure and the second measure.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of Mummadi in view of Xu and Goswami before them, to include Oberbreckling’s determining a weighted sum of the first and second measure. The motivation for doing so would have been to improve the processing of a computer system (See[0036] of Oberbreckling ).
Claims 15 is rejected under 35 U.S.C 103 as being unpatentable over Mummadi in view of Xu, Goswami and further in view of Garera et al. (Pub No.: 20140314311 A1), hereinafter referred to as Garera.
Regarding claim 15, Mummadi in view of Xu and Goswami disclose elements of claim 14. Mummadi in view of Xu and Goswami does not explicitly disclose:
The method according to claim 14, wherein the sensing system includes one or more output devices and one or more input devices, and wherein the method further comprises: - outputting via an output device a request for a user to approve or disapprove a determined classification
However, Garera disclose the limitation (In paragraph [0060], Gare discloses generating classification using a machine-learning classification model and presenting classification to a user for manual validation. Permitting the analyst to approve or disapprove the classification decision and/or provide an alternative classification.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of Mummadi in view of Xu and Goswami before them, to include Garera with transmitting some or all of the validations decisions to one or more analyst workstations as taught by Garera. The motivation for doing so would have been to improve methods for incorporating both automated classification and human judgment in a highly effective manner (See [0006] of Garera.)
Allowable Subject Matter
Claim 10 is objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Response to Arguments
Applicant's arguments filed 08/19/2026 have been fully considered, but in part are not persuasive.
Pertaining to rejection under 101
Rejections for claims 1-11 and 14-17 are withdrawn under 35 USC § 101
Pertaining to rejection under 103
Thus, the applicant’s arguments in regard to the examiner’s rejections under 35 USC 103 are moot in view of the new grounds of rejection.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to EVEL HONORE whose telephone number is (703)756-1179. The examiner can normally be reached Monday-Friday 8 a.m. -5:30 p.m.
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, Mariela D Reyes can be reached at (571) 270-1006. 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.
EVEL HONORE
Examiner
Art Unit 2142
/Mariela Reyes/Supervisory Patent Examiner, Art Unit 2142