Indirect Invisible Poisoning Attacks on Domain Adaptation

Abstract

Unsupervised domain adaptation has been successfully applied across multiple high-impact applications, since it improves the generalization performance of a learning algorithm when the source and target domains are related. However, the adversarial vulnerability of domain adaptation models has largely been neglected. Most existing unsupervised domain adaptation algorithms might be easily fooled by an adversary, resulting in deteriorated prediction performance on the target domain, when transferring the knowledge from a maliciously manipulated source domain. To demonstrate the adversarial vulnerability of existing domain adaptation techniques, in this paper, we propose a generic data poisoning attack framework named I2Attack for domain adaptation with the following properties: (1) perceptibly unnoticeable: all the poisoned inputs are natural-looking; (2)adversarially indirect: only source examples are maliciously manipulated; (3) algorithmically invisible: both source classification error and marginal domain discrepancy between source and target domains will not increase. Specifically, it aims to degrade the overall prediction performance on the target domain by maximizing the label-informed domain discrepancy over both input feature space and class-label space be-tween source and target domains. Within this framework, a family of practical poisoning attacks are presented to fool the existing domain adaptation algorithms associated with different discrepancy measures. Extensive experiments on various domain adaptation benchmarks confirm the effectiveness and computational efficiency of our proposed I2Attack framework.