Defending Against Multiple and Unforeseen Adversarial Videos

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) face adversarial attacks, but current defenses often lack multi-perturbation robustness.
• Adversarial robustness in video recognition is less explored than in image recognition, with limited defense strategies available.

Purpose of the Study:

• To propose one of the first defense strategies against multiple types of adversarial videos for video recognition.
• To enhance the multi-perturbation robustness of video recognition models against diverse adversarial attacks.

Main Methods:

• Introduced MultiBN, a defense method using adversarial training with multiple independent batch normalization (BN) layers.
• Incorporated a learning-based BN selection module to automatically detect attack types and route videos to appropriate BN branches.
• Enabled end-to-end training for a fully automatic defense system.

Main Results:

• MultiBN demonstrated stronger multi-perturbation robustness against various adversarial video types, including Lp-bounded and physically realizable attacks.
• The proposed method showed effectiveness across different datasets and target models, outperforming existing adversarial training approaches.
• Analysis confirmed the benefits of the multiple BN structure for handling diverse perturbation types.

Conclusions:

• MultiBN offers a significant advancement in defending video recognition models against multiple adversarial video perturbations.
• The approach provides a robust and automatic solution for enhancing adversarial robustness in video recognition systems.
• This work opens new avenues for research in multi-perturbation adversarial defense for video analysis.