arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2209.07125cs.CR
收藏

BadRes:通过残余连接揭露后门

BadRes: Reveal the Backdoors through Residual Connection

Mingrui He, Tianyu Chen, Haoyi Zhou, Shanghang Zhang, Jianxin Li

摘要:

通常,在为CV和VL中的各种下游任务构建CNN和Transformers时,剩余连接是不可或缺的网络组件,这鼓励跳过网络块之间的快捷方式。然而,逐层冗余连接也可能通过允许不知情的输入而损害模型的鲁棒性。本文提出了一种简单而强大的后门攻击方法- BadRes,其中剩余连接充当旋转门,在干净输入上是确定性的,而在中毒输入上是不可预测的。我们使用ViT和BEiT模型对四个数据集进行了实证评估,BadRes实现了97%的攻击成功率,同时在干净数据上获得零性能下降。此外,我们分析了BadRes与国家的最先进的防御方法,揭示了根本的弱点在于剩余连接。

英文摘要:

Generally, residual connections are indispensable network components in building CNNs and Transformers for various downstream tasks in CV and VL, which encourages skip shortcuts between network blocks. However, the layer-by-layer loopback residual connections may also hurt the model's robustness by allowing unsuspecting input. In this paper, we proposed a simple yet strong backdoor attack method - BadRes, where the residual connections play as a turnstile to be deterministic on clean inputs while unpredictable on poisoned ones. We have performed empirical evaluations on four datasets with ViT and BEiT models, and the BadRes achieves 97% attack success rate while receiving zero performance degradation on clean data. Moreover, we analyze BadRes with state-of-the-art defense methods and reveal the fundamental weakness lying in residual connections.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑