With the development of Deep Neural Network (DNN), as well as the demand growth of third-party DNN model stronger, there leaves a gap for backdoor attack. Backdoor can be injected into a third-party model and has strong stealthiness in normal situation, thus has been widely discussed. Nowadays backdoor attack on deep neural network has been concerned a lot and there comes lots of researches about attack and defense around backdoor in DNN. In this paper, we propose a robust avatar backdoor attack that integrated with adversarial attack. Our attack can escape mainstream detection schemes with popularity and impact that detect whether a model has backdoor or not before deployed. It reveals that although many effective backdoor defense schemes has been put forward, backdoor attack in DNN still needs to be concerned. We select three popular datasets and two detection schemes with high impact factor to prove that our attack has a great performance in aggressivity and stealthiness.
翻译:随着深神经网络(DNN)的发展,以及第三方DNN模式需求的增长,后门攻击留下一个缺口。后门可以注入第三方模式,在正常情况下具有很强的隐秘性,因此已经对此进行了广泛讨论。如今,深神经网络的后门攻击引起了很大的关注,对DNN的后门攻击也产生了大量关于后门攻击和防御的研究。在这份文件中,我们提出了与对抗性攻击相结合的强健的后门攻击。我们的攻击可以逃脱主流的探测计划,其受欢迎和影响可以探测出模型是否在后门部署之前是否已经部署过。它表明,虽然许多有效的后门防御计划已经提出,但DNNN的后门攻击仍然需要关注。我们选择了三个广受欢迎的数据集和两个具有高度影响力的探测计划,以证明我们的攻击在侵略性和隐秘性方面表现良好。