Implementing a Multitarget Backdoor Attack Algorithm Based on Procedural Noise Texture Features

Abstract

Recent studies have shown that deep neural networks (DNNs) may suffer from some security issues such as backdoor attacks. The triggers of backdoor attacks are dynamic and global. However, existing backdoor attack methods are laborious and time-consuming in hiding global triggers, and most of them focus on single-target attacks, with less research on multitarget backdoor attacks. In this work, we present a multitarget attack strategy using texture features of procedural noise. Specifically, we use the k-LSB steganography algorithm to hide the triggers in the image and use different texture features of procedural noise to trigger multiple targets for attack. Poisoned images can be generated more quickly using the k-LSB steganography algorithm without any training process. Multitarget backdoor attacks apply to more scenarios and are more difficult to defend against. We evaluate the effectiveness of the proposed attack method on GTSRB and ImageNet datasets, and the experiments show that the proposed attack can achieve a high attack success rate (up to 100.00% for GTSRB and up to 98.48% for ImageNet) without compromising the clean data’s categorization performance, and thus is less likely to arouse the suspicion of administrators. In addition, the attack can bypass existing defense methods (STRIP defense and Neural Cleanse defense).

Keywords

• Backdoor attack
• deep learning
• procedural noise
• LSB steganography