Phase-Based Invisible Backdoor Attack

Le résumé fourni par la source

Deep Neural Networks (DNNs) are vulnerable to backdoor attacks. However, most backdoor attacks produce poisoned images with detectable spatial or frequency artifacts, and even invisible attacks often lack spatial smoothness or color consistency. To address these limitations, we propose PIBA, a novel phase-based invisible backdoor attack. PIBA employs Discrete Wavelet Transform (DWT) to extract low-scale components, followed by color transformation and slide-window-based Fast Fourier Transform (FFT) to embed phase triggers, ensuring robust attack performance without spatial or frequency artifacts. We evaluate PIBA on four image datasets and two model architectures, comparing it with five state-of-the-art backdoor attacks and testing its robustness against two advanced defenses. Experimental results show that PIBA achieves attack performance on par with leading methods while delivering state-of-the-art stealthiness across all datasets.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.