Perceptually Imperceptible Backdoor Attacks Using High-Frequency Information in Deep Learning Models

Backdoor attacks, which involve injecting malicious triggers into training data, pose a significant threat to the security of deep learning models across various domains. This paper introduces a trigger injection method that capitalises on the model's high-frequency feature learning ability. The method uses imperceptible high-frequency features to conceal the backdoor during training and accumulates high-frequency components as triggers for a covert clean-label backdoor attack in the inference phase. This approach ensures high concealment while preserving image quality and perception. It highlights the broader significance of understanding human visual systems for enhancing the reliability of deep learning models in diverse applications.