Backdoor Attack Based on Lossy Image Compression Using Discrete Cosine Transform

Deep neural networks (DNNs) have been widely used in the field of image recognition. The advent of image backdoor attacks poses significant security threats to the use of DNNs. Researching advanced backdoor attacks is a prerequisite for developing defense methods that enhance the security of DNNs. However, existing backdoor attacks face challenges in simultaneously achieving invisibility, effectiveness, and time efficiency. Therefore, for image classification tasks based on DNNs, we propose a novel backdoor attack based on lossy image compression using discrete cosine transform (DCTLICBA). This method uses the lossy compression technique to introduce noise that is imperceptible to human vision as a backdoor trigger. Then, an invisible, effective, and time-efficient backdoor attack is launched through data poisoning. We verified our proposed method on the ImageNet dataset. Experimental results indicate that DCTLICBA achieves a high attack success rate while maintaining visual invisibility and time efficiency. Furthermore, our proposed method can bypass typical backdoor defense methods.