US 12,008,079 B1
Process to make machine object detection robust to adversarial attacks
Soheil Kolouri, Agoura Hills, CA (US); Heiko Hoffmann, Simi Valley, CA (US); and David W. Payton, Calabasas, CA (US)
Assigned to HRL LABORATORIES, LLC, Malibu, CA (US)
Filed by HRL Laboratories, LLC, Malibu, CA (US)
Filed on Jul. 6, 2021, as Appl. No. 17/368,635.
Claims priority of provisional application 63/060,494, filed on Aug. 3, 2020.
Int. Cl. G06F 18/24 (2023.01); G06F 18/21 (2023.01); G06F 18/214 (2023.01); G06F 18/232 (2023.01); G06F 21/56 (2013.01); G06N 3/08 (2023.01); G06V 10/22 (2022.01)
CPC G06F 18/24 (2023.01) [G06F 18/214 (2023.01); G06F 18/2163 (2023.01); G06F 18/217 (2023.01); G06F 18/232 (2023.01); G06F 21/566 (2013.01); G06N 3/08 (2013.01); G06V 10/22 (2022.01); G06F 2221/034 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A system for object detection that is robust to adversarial attacks, the system comprising:
one or more processors and a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors perform operations of:
generating an initial hypothesis of an identity of an object in an input image using a sparse convolutional neural network (CNN) and a distribution aware classifier;
performing a foveated hypothesis verification process, wherein performing the foveated hypothesis verification process comprises identifying a region of the input image that supports the initial hypothesis;
using a part-based classifier, predicting an identity of a part of the object in the region of the input image;
determining an attack probability for the predicted identity of the part;
updating the initial hypothesis based on the predicted identity of the part and the attack probability;
performing the foveated hypothesis verification process and updating of hypotheses until a hypothesis reaches a certainty threshold;
labeling the object based on the hypothesis that reached the certainty threshold; and
controlling an action performed by an autonomous platform based on the labeling of the object.