Toward Enhanced Adversarial Robustness Generalization in Object Detection: Feature Disentangled Domain Adaptation for Adversarial Training

초록

Recent research has shown that deep learning models are likely to make incorrect predictions even when exposed to minor perturbations. To address this, training models on adversarial examples, particularly through Adversarial Training (AT), has gained attraction. However, traditional AT is prone to overfitting to specific attack types and remains vulnerable to other kinds of attacks. To solve this problem, we propose Feature Disentangled Domain Adaptation (FDDA). FDDA enhances the robustness of deep learning models through domain adaptation, separating the features of clean and adversarial images. Additionally, by introducing Feature Recalibration, the proposed method ensures more consistent learning of shared features between the two domains. Experimental results show FDDA's effectiveness against different adversarial attacks compared to traditional methods. By minimizing conflicts between clean and adversarial images, FDDA maximizes clean accuracy, demonstrating its superiority over state-of-the-art approaches.

키워드