CubeAgent: Efficient query-based video adversarial examples generation through deep reinforcement learning

Abstract

In commercial deep-learning-based video systems, testers utilize query-based methods to generate adversarial examples (AEs) and effectively uncover system vulnerabilities. The current research has primarily focused on selecting the key perturbation units, such as video patches, keyframes, and combinations of keyframes and regions, to add adversarial perturbation and generate AEs. Furthermore, deep reinforcement learning (DRL) frameworks have been utilized to model the results of sequence-based feedback to reduce query numbers. However, considering the pixels of spatial and temporal dimensions separately in the search process results in a large number of queries and intolerable failure rates for video AE generation. This paper proposes a new AEs perturbation unit called the “video cube”, which simultaneously extracts video pixels in neighbor frames and regions. We develop a new DRL framework called “CubeAgent”, which incorporates controllable policy actions for selecting the number and index of key video cubes segmented by time intervals. We conducted exhaustive experiments across diverse video DNN systems, utilizing the UCF101 and JESTER datasets, which conclusively demonstrated that CubeAgent can expedite the generation process by a factor of two, diminishing the average query count from 5,768 to 4,602, representing a 20% reduction, while simultaneously mitigating the average generation failure rate from 9% to 7%. The results show that CubeAgent improves the performance of adversarial example generation while achieving comparable.