Efficient Robustness Assessment via Adversarial Spatial-Temporal Focus on Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning Security

Background:

• Video recognition models are crucial for safety-critical applications.
• High dimensionality of videos poses computational challenges for adversarial attacks, especially black-box scenarios.
• Existing methods struggle with efficiency due to extensive query requirements.

Purpose of the Study:

• To develop an efficient and effective method for adversarial robustness assessment in video recognition.
• To reduce the computational cost and query number in black-box adversarial attacks on videos.
• To propose a novel attack strategy that mitigates redundancy in video data.

Main Methods:

• Introduced the Adversarial spatial-temporal Focus (AstFocus) attack.
• Utilized a cooperative Multi-Agent Reinforcement Learning (MARL) framework with agents for key frame and region selection.
• Jointly trained agents using rewards from black-box threat models to optimize cooperative prediction.
• Reduced the search space by focusing on spatio-temporal redundancies.

Main Results:

• AstFocus significantly reduces the number of queries needed for gradient estimation.
• The attack demonstrates superior performance across four mainstream video recognition models and three action recognition datasets.
• Achieved improved fooling rates, reduced query numbers, faster execution times, and smaller perturbation magnitudes compared to state-of-the-art methods.
• Effectively mitigates both temporal and spatial redundancy.

Conclusions:

• AstFocus offers a computationally efficient and effective solution for adversarial robustness assessment in video recognition.
• The MARL-based approach successfully targets key spatio-temporal elements, optimizing attack performance.
• This method provides a significant advancement in defending video recognition systems against adversarial attacks.