Sparse Adversarial Video Attacks via Superpixel-Based Jacobian Computation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial examples, which pose a significant threat.
• Adversarial attacks have been extended to video models, with sparse perturbations offering advantages in computation and stealth.

Purpose of the Study:

• To propose an efficient attack method for generating highly sparse adversarial perturbations in both temporal and spatial domains of videos.
• To address the challenge of dimensional explosion in video analysis using super-pixels.

Main Methods:

• Utilized forward derivative feedback to identify key frames and pixels for perturbation.
• Introduced super-pixels to reduce the computational complexity of gradient calculations.
• Validated the method in both white-box and black-box attack scenarios, using Natural Evolution Strategy (NES) for gradient estimation in black-box settings.

Main Results:

• Achieved comparable attack performance to state-of-the-art methods.
• Polluted less than 1% of pixels, demonstrating high sparsity.
• Required less time to complete attacks compared to existing techniques.

Conclusions:

• The proposed method offers an efficient and stealthy approach to generate adversarial video perturbations.
• The use of super-pixels effectively mitigates the dimensional explosion problem in video adversarial attacks.
• This research contributes to understanding and defending against adversarial attacks on video deep learning models.