Defense Contrastive Poisoning: An Application of JPEG to Self-Supervised Contrastive Learning Indiscriminate Poisoning Attacks

Indiscriminate poisoning attacks are a particular type of data poisoning attack in which the attacker adds perturbations to the training data or labels to interfere with the learning process of the model, thus affecting its performance and usability. Self-supervised learning is a new artificial intelligence paradigm that does not require labeling of the dataset during training. Contrastive Poisoning is a method of Indiscriminate poisoning attacks for self-supervised contrastive learning. Contrastive Poisoning has shown excellent attack results on several self-supervised learning algorithms. In this paper, we propose a JPEG-based method, Noise-JPEG, for defense against Contrastive Poisoning. We test the effectiveness of Noise-JPEG’s defense against several self-supervised contrastive learning algorithms on CIFAR-10 and CIFAR-100 datasets. The results show that our method is effective on different datasets and algorithms. Exhibit stable and effective defense performance. Noise-JPEG outperforms other previously studied countermeasures, such as adversarial training and matrix completion. Our method will increase the accuracy of models attacked by Contrastive Poisoning from 44.9% to 86.4%.