US 12,321,856 B2
Method, computer program and device for evaluating the robustness of a neural network against image disturbances
Alfred Laugros, Grenoble (FR)
Assigned to BULL SAS, Les Clayes-sous-Bois (FR)
Filed by BULL SAS, Les Clayes-sous-bois (FR)
Filed on Jun. 2, 2021, as Appl. No. 17/336,507.
Claims priority of application No. 20305577 (EP), filed on Jun. 2, 2020.
Prior Publication US 2021/0374482 A1, Dec. 2, 2021
Int. Cl. G06N 3/08 (2023.01); G06F 18/21 (2023.01); G06F 18/22 (2023.01); G06V 10/764 (2022.01)
CPC G06N 3/08 (2013.01) [G06F 18/217 (2023.01); G06F 18/22 (2023.01); G06V 10/764 (2022.01)] 17 Claims
OG exemplary drawing
 
1. A method for evaluating robustness of a neural network used for image processing against a group of at least two different image disturbances, said method comprising:
for each image disturbance of said group of said at least two different image disturbances, determining an individual robustness score of said neural network, representative of a robustness of the neural network against said each image disturbance;
for at least one pair of image disturbances of said group, determining a corresponding individual similarity score, representative of a similarity between the at least two different image disturbances of said at least one pair of image disturbances; and
calculating an overall robustness score of said neural network against said group of at least two different image disturbances, as a function of:
each determined individual robustness score, and
the corresponding individual similarity score determined for the at least one pair of image disturbances;
wherein the corresponding individual similarity score for the similarity of a first image disturbance of the at least one pair of image disturbances to a second image disturbance of the at least one pair of image disturbances is equal to the corresponding individual similarity score for the similarity of the second image disturbance to the first image disturbance.