Improving Generalization of Adversarial Training via Robust Critical Fine Tuning | |
Zhu, Kaijie3,4; Hu, Xixu2; Wang, Jindong1; Xie, Xing1; Yang, Ge3,4 | |
2023 | |
会议日期 | 2023-9 |
会议地点 | Paris, France |
英文摘要 | Deep neural networks are susceptible to adversarial ex- amples, posing a significant security risk in critical applica- tions. Adversarial Training (AT) is a well-established tech- nique to enhance adversarial robustness, but it often comes at the cost of decreased generalization ability. This paper proposes Robustness Critical Fine-Tuning (RiFT), a novel approach to enhance generalization without compromising adversarial robustness. The core idea of RiFT is to exploit the redundant capacity for robustness by fine-tuning the ad- versarially trained model on its non-robust-critical module. To do so, we introduce module robust criticality (MRC), a measure that evaluates the significance of a given mod- ule to model robustness under worst-case weight perturba- tions. Using this measure, we identify the module with the lowest MRC value as the non-robust-critical module and fine-tune its weights to obtain fine-tuned weights. Subse- quently, we linearly interpolate between the adversarially trained weights and fine-tuned weights to derive the optimal fine-tuned model weights. We demonstrate the efficacy of RiFT on ResNet18, ResNet34, and WideResNet34-10 mod- els trained on CIFAR10, CIFAR100, and Tiny-ImageNet datasets. Our experiments show that RiFT can significantly improve both generalization and out-of-distribution robust- ness by around 1.5% while maintaining or even slightly enhancing adversarial robustness. Code is available at https://github.com/Immortalise/RiFT. |
语种 | 英语 |
内容类型 | 会议论文 |
源URL | [http://ir.ia.ac.cn/handle/173211/56687] ![]() |
专题 | 模式识别国家重点实验室_计算生物学与机器智能 |
通讯作者 | Yang, Ge |
作者单位 | 1.Microsoft Research 2.City University of Hong Kong 3.School of Artificial Intelligence, University of Chinese Academy of Sciences 4.Institute of Automation, Chinese Academy of Sciences |
推荐引用方式 GB/T 7714 | Zhu, Kaijie,Hu, Xixu,Wang, Jindong,et al. Improving Generalization of Adversarial Training via Robust Critical Fine Tuning[C]. 见:. Paris, France. 2023-9. |
个性服务 |
查看访问统计 |
相关权益政策 |
暂无数据 |
收藏/分享 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论