Abstract
Adversarial purification offers a flexible defense against unseen attacks in person re-identification (re-ID) systems. However, existing methods often suffer from indiscriminate purification that degrades benign images, fixed intensity regardless of attack strength, and high computational overhead. We present IntensPure, a framework that adaptively purifies images by estimating attack intensity and operating within a reduced frequency domain. By focusing on the most vulnerable low-frequency components of the Block Discrete Cosine Transform (BDCT), IntensPure achieves state-of-the-art accuracy while significantly improving inference speed.
Introduction
Person re-identification (re-ID) is critical for urban safety and multi-camera tracking. Despite its importance, re-ID models are highly susceptible to adversarial attacks—carefully crafted perturbations that lead to misidentification. While adversarial training is a common defense, it lacks generalization to novel attack types. Adversarial purification, which uses generative models like diffusion to remove noise, is a promising alternative.
Current purification techniques face three primary challenges: they often degrade clean (benign) images, use a "one-size-fits-all" approach to attack intensity, and are computationally expensive due to processing high-resolution spatial data. IntensPure addresses these by introducing an intensity-aware, frequency-focused diffusion process.
Methodology
The IntensPure framework consists of four main modules designed to optimize both defense effectiveness and computational efficiency.
- Attack Intensity Estimation: The system first estimates the attack strength ($\epsilon$) by analyzing ID Stability (consistency of top-10 retrieval results) and Attribute Inconsistency (errors in predicted traits like clothing color) through an auxiliary network.
- Secondary Image Generation: Instead of diffusing the entire spatial image, IntensPure applies an 8x8 Block Discrete Cosine Transform (BDCT). It isolates the six lowest-frequency coefficients, which are empirically proven to be the most vulnerable to re-ID attacks, reducing the data dimension by approximately 90%.
- Secondary Domain Adaptive Diffusion: Using the estimated intensity, the model determines an optimal diffusion time step ($t^*$) via a logarithmic relationship. It then performs directional diffusion on the selected coefficients using specialized convolutional layers.
- Spatial Image Reconstruction: The purified coefficients are combined with the original high-frequency components and transformed back to the spatial domain via inverse BDCT for final feature extraction.
Experimental Results
IntensPure was evaluated on the Market1501 and DukeMTMC-reID benchmarks against Metric-FGSM, Deep Mis-Ranking, and MetaAttack.
Performance and Robustness
Under a strong Metric-FGSM attack ($\epsilon=16$) on Market1501, IntensPure maintained a 60.42% rank-1 accuracy, significantly outperforming other state-of-the-art methods like GNSP (54.44%). Crucially, on benign images, IntensPure achieved 88.36% accuracy, nearly matching the 88.84% baseline. In contrast, traditional methods like DiffPure dropped benign accuracy to approximately 74.73% due to over-purification.
Efficiency
By operating in the frequency domain and using adaptive steps, IntensPure reduced inference time from 366ms (DiffPure) to 64ms. Computational requirements were also slashed from 583G FLOPs to 50G FLOPs, making it viable for real-time applications.
Key Insights
- Frequency Sensitivity: Re-ID models are primarily affected by low-frequency adversarial noise. Focusing purification on just 6 out of 64 BDCT coefficients allows for effective defense with minimal data processing.
- Adaptive Defense: The logarithmic fitting of the relationship between attack intensity and optimal diffusion time ensures that the defense is neither too weak to remove noise nor too strong to destroy identity.
Conclusion
IntensPure represents a significant step forward in adversarial defense for re-ID. By combining attack intensity estimation with frequency-domain diffusion, it provides a robust, efficient, and identity-preserving purification process. Future work will explore the system's resilience in white-box scenarios and its applicability to other fine-grained recognition tasks.
Summary updated