CLPA: Clean-Label Poisoning Availability Attacks Using Generative Adversarial Nets

Poisoning attacks are emerging threats to deep neural networks where the adversaries attempt compromise models by injecting malicious data points in clean training data. target either availability or integrity of a model. The attack aims degrade overall accuracy while causes misclassification only for specific instances without affecting Although clean-label proven be effective recent studies, feasibility remains unclear. This paper, first time, proposes approach, CLPA, poisoning attack. We reveal that due intrinsic imperfection classifiers, naturally misclassified inputs can considered as special type poisoned data, which we refer "natural data''. then propose two-phase generative adversarial net (GAN) based generation framework along with triplet loss function synthesizing samples locate similar distribution natural generated plausible human perception and also bypass singular vector decomposition (SVD) defense. demonstrate effectiveness our approach on CIFAR-10 ImageNet dataset over variety models. Codes available at: https://github.com/bxz9200/CLPA.