Efficient robustness assessment via adversarial spatial-temporal focus on videos

Songping Wang

Abstract: Adversarial robustness assessment for video recognition models has raised concerns owing to their wide applications on safety-critical tasks like security surveillance. Compared with images, videos have much high dimension because of the additional temporal information, which brings high computational costs when generating adversarial videos. This is especially significant for the query-based black-box attacks because they usually need gradient estimation for the threat models, and high dimensions lead to a large number of queries. Previous methods mitigate this issue by mainly reducing the temporal redundancy between frames. In this paper, we argue that eliminating spatial redundancy within frames can also contribute to the video's gradient estimation as done in the image, and more importantly, this inspires us to jointly eliminate the temporal and spatial redundancy to achieve an effective and efficient gradient estimation on the reduced searching space, and thus query number could decrease. To implement this idea, we design the novel Adversarial spatial-temporal Focus (AstFocus) attack on videos, which performs attacks on the simultaneously focused key frames and key regions from the inter-frames and intra-frames in the video. AstFocus attack is based on the cooperative Multi-Agent Reinforcement Learning (MARL) framework. One agent is responsible for selecting key frames, and another agent is responsible for selecting key regions. These two agents are jointly trained by the common rewards received from the black-box threat models to perform a cooperative prediction. By continuously querying, the reduced searching space composed of key frames and key regions is becoming precise, and the whole query number becomes less than that on the original video. Extensive experiments on two mainstream video recognition models and two widely used action recognition datasets demonstrate that the proposed AstFocus attack outperforms the SOTA methods, which is prevenient in fooling rate, query number, and perturbation magnitude at the same.