Real-Time Adversarial Purification via DINOv2-Guided Contrastive Embedding Alignment

Abstract

Deep neural networks are vulnerable to misclassification even with small adversarial noise. Although purification methods using diffusion models have improved defense performance, they suffer from inference latency due to multiple sampling steps, making real-time applications difficult. In this paper, targeting a practical black-box purification scenario, we propose a feed-forward purification model composed of four ResBlocks, called DINOv2-guided Contrastive Embedding Alignment (CEA), to reduce this computation delay. The proposed model employs adversarial training by generating AutoAttack examples inside the training loop. To alleviate visual distortions of original images during the purification process, we introduce contrastive learning using a pre-trained DINOv2 as a teacher model. Experimental results show that the proposed model achieved an inference speed of 0.0019 seconds per image with a mere 0.03 TFLOPs, and recorded a robust accuracy of 70.76% on the ResNet-50 target model. This indicates that the proposed method effectively mitigates the latency problem of diffusion-based defense techniques and provides the robustness required for real-time processing environments. Consequently, by requiring only 0.03 TFLOPs, this efficient purification mechanism shows potential to secure real-time AI services on edge computing devices with limited computing resources, providing a highly practical defense for mission-critical M3 (Mobile, Military, Maritime) operations against instantaneous adversarial threats.