Anti-Interpolation: An Attack Facilitator Hiding Adversaries Into Images

Deep learning models show the vulnerability to adversarial examples, which can be attacked to make incorrect predictions. As one of the key image preprocessing techniques, interpolation may weaken the robustness of the attack and even disable the perturbation. Thus, the malicious images cannot achieve the attack purpose. In this work, anti-interpolation, as an attack facilitator, is proposed, where interpolation does not affect the attack results intended by the perturbation. Based on anti-interpolation, an attack framework is extended to achieve adversarial attacks, including targeted attack, non-targeted attack, and auxiliary attack. Furthermore, gradient derivation is developed based on the interpolation output of large-sized images, which can infer the interpolation mode in the black box attack scenario. Compared with the current adversarial attacks, anti-interpolation can achieve a higher attack success rate. Moreover, anti-interpolation is capable of attacking defensive models with various model corrections.