Temporal shuffling for defending deep action recognition models against adversarial attacks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) excel in video-based action recognition.
• The generalization mechanisms of these models remain poorly understood.
• Existing models may not fully leverage motion dynamics.

Purpose of the Study:

• Investigate the generalization mechanisms of action recognition models.
• Identify vulnerabilities to adversarial attacks.
• Develop a novel defense strategy against such attacks.

Main Methods:

• Analyzing the robustness of action recognition models to frame order randomization.
• Examining the role of motion information and monotonicity.
• Proposing a defense method based on temporal shuffling of video frames.

Main Results:

• Action recognition models show unexpected robustness to frame order randomization.
• Motion information is utilized less than anticipated by these models.
• Adversarial perturbations are sensitive to temporal disruptions in videos.

Conclusions:

• Action recognition models' reliance on motion is less than expected, contributing to robustness.
• Temporal shuffling is an effective defense against adversarial attacks for 3D CNNs.
• This defense method requires no additional training, offering a practical solution.