Deep neural networks (DNNs) can be manipulated to exhibit specific behaviors when exposed to specific trigger patterns, without affecting their performance on benign samples, dubbed backdoor attack. Some recent research has focused on designing invisible triggers for backdoor attacks to ensure visual stealthiness, while showing high effectiveness, even under backdoor defense. However, we find that these carefully designed invisible triggers are often sensitive to visual distortion during inference, such as Gaussian blurring or environmental variations in physical scenarios. This phenomenon could significantly undermine the practical effectiveness of attacks, but has been rarely paid attention to and thoroughly investigated. To address this limitation, we define a novel trigger called the Visible, Semantic, Sample-Specific, and Compatible trigger (VSSC trigger), to achieve effective, stealthy and robust to visual distortion simultaneously. To implement it, we develop an innovative approach by utilizing the powerful capabilities of large language models for choosing the suitable trigger and text-guided image editing techniques for generating the poisoned image with the trigger. Extensive experimental results and analysis validate the effectiveness, stealthiness and robustness of the VSSC trigger. It demonstrates superior robustness to distortions compared with most digital backdoor attacks and allows more efficient and flexible trigger integration compared to physical backdoor attacks. We hope that the proposed VSSC trigger and implementation approach could inspire future studies on designing more practical triggers in backdoor attacks.

翻译：暂无翻译