Poster Abstract: Leveraging Phase Offset for Stealthy Backdoor Attacks

Deep neural networks (DNNs) are increasingly exposed to backdoor attacks during model training, particularly when conducted through third-party services. This work introduces a novel backdoor strategy that employs phase offsets within various frequency bands to improve the anti-interference capabilities of trigger mechanisms in image classification models. By adjusting the phase components selectively, our method embeds triggers that are difficult to detect, subtly manipulating model responses while preserving image quality. Finally, the experimental results show that our attack method is effective.