Towards Effective and Efficient Adversarial Defense with Diffusion Models for Robust Visual Tracking

2025-05-31

AI Summary

Visual tracking models exhibit insufficient robustness against adversarial attacks, leading to significant performance degradation. This paper proposes the first zero-shot, real-time defense framework leveraging a lightweight diffusion prior—requiring no adversarial training—to enhance perturbation resilience. Our method integrates the diffusion prior into a Siamese tracking architecture to perform adversarial purification directly in feature space. We further introduce a noise-aware feature calibration mechanism that jointly optimizes robustness and efficiency, and refine the noise scheduling strategy to minimize computational overhead. Evaluated on LaSOT and GOT-10k benchmarks under diverse adversarial attacks, our approach achieves a 23.6% improvement in mean Average Precision (mAP) while incurring only a 1.8 ms increase in inference latency—substantially outperforming existing defense methods in both accuracy and efficiency.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Adversarial Learning & Robustness
• Search and Optimization: Adversarial Search

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Tracking, profiling, and countermeasures against them
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Problem

• Improving robustness of visual tracking against adversarial attacks
• Proposing diffusion-based defense method for real-time performance
• Combining multi-scale losses to suppress adversarial perturbations

Innovation

• Denoise diffusion models for adversarial defense
• Multi-scale loss mechanism suppresses perturbations
• Real-time robust tracking at 30 FPS