Cambackdoor: a self-supervised backdoor attack method against CLP pruning defense

Self-supervised learning(SSL) has made significant progress in representation learning from unlabeled data but has also exposed vulnerabilities to backdoor attacks. Existing backdoor attack methods for self-supervised learning demonstrate strong attack effectiveness; however, their performance significantly degrades when defenders apply the Channel Lipschitz Pruning (CLP) defense strategy. To address this issue, this paper proposes a novel backdoor attack method, CamBackdoor, which is specifically designed to be adversarial against CLP defense. Based on an in-depth study of the CLP pruning method and the principles of SSL, CamBackdoor strategically selects specific channels within the model for backdoor injection. After the initial backdoor injection, it further Camouflages the backdoor channels, enhancing their stealthiness under CLP pruning defenses. This enables more effective backdoor attacks in CLP defense scenarios. We evaluate the effectiveness of CamBackdoor under CLP defense using two different SSL models, SimCLR and BYOL, on the CIFAR-10 and ImageNet-100 datasets. Additionally, we compare CamBackdoor with three representative backdoor attack methods: CTRL, BadEncoder, and ESTAS. Experimental results strongly demonstrate that CamBackdoor exhibits significant advantages in CLP-pruned defense scenarios.