Adversarial training, as an important technique for enhancing model robustness, faces problems of high training costs and inability to defend against multiple adversarial attacks. An improved adversarial training approach based on meta-learning is proposed. By integrating pre-training fine-tuning and diffusion model data generation strategies, a dual-branch training architecture is designed. One branch is fine-tuned on an l∞ robust model to improve its l∞ robustness, and the other branch trains against composite adversarialattacks to enhance the model’s defense capabilities against non-lp norm attacks. During training, the weights of both branches are fused through a mixed model and periodically reinitialized, enabling the final model to simultaneously resist both l∞ attacks and composite adversarial attacks. Experimental results show that the proposed approach maintains l∞ robustness while achieving superior defensive performance against composite adversarial attacks on the composite adversarial robustness benchmark (CARBEN).
MADRYA, MAKELOVA, SCHMIDTL, et al. Towards deep learning models resistant to adversarial attacks[DB/OL]. (2019-09-04)[2024-11-05].
[3]
ZHANGH Y, YUY D, JIAOJ T, et al. Theoretically principled trade-off between robustness and accuracy[C]∥Proceedings of the 36th International Conference on Machine Learning. New York, USA: PMLR, 2019:7472-7482.
[4]
WANGZ K, PANGT Y, DUC, et al. Better diffusion models further improve adversarial training[C]∥Proceedings of the 40th International Conference on Machine Learning. New York, USA: PMLR, 2023:36246-36263.
[5]
HSIUNGL, TSAIY Y, CHENP Y, et al. Towards compositional adversarial robustness: generalizing adversarial training to composite semantic perturbations[C]∥Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2023:24658-24667.
[6]
LAIDLAWC, FEIZIS. Functional adversarial attacks[C]∥Proceedings of the 33rd International Conference on Neural Information Processing Systems. Red Hook, USA: Curran Associates Inc., 2019:10408-10418.
[7]
CROCEF, ANDRIUSHCHENKOM, SEHWAGV, et al. RobustBench: a standardized adversarial robustness benchmark[DB/OL]. (2021-10-11)[2024-11-05].
[8]
WUD X, XIAS T, WANGY S. Adversarial weight perturbation helps robust generalization[C]∥Proceedings of the 34th International Conference on Neural Information Processing Systems. New York, USA: ACM, 2020:2958-2969.
[9]
MADAAND, SHINJ, HWANGS J. Learning to generate noise for multi-attack robustness[C]∥Proceedings of the 38th International Conference on Machine Learning. New York, USA: PMLR, 2021:7279-7289.
[10]
LIUZ Q, XUY, JIX Y, et al. TWINS: a fine-tuning framework for improved transferability of adversarial robustness and generalization[C]∥Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2023:16436-16446.
[11]
ZHUK J, HUX X, WANGJ D, et al. Improving generalization of adversarial training via robust critical fine-tuning[C]∥Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway, USA: IEEE, 2023:4401-4411.
[12]
LEEY, CHENA S, TAJWARF, et al. Surgical fine-tuning improves adaptation to distribution shifts[DB/OL]. (2023-06-06)[2024-11-05].
[13]
CROCEF, HEINM. Adversarial robustness against multiple and single lp -threat models via quick fine-tuning of robust classifiers[C]∥Proceedings of the 39th International Conference on Machine Learning. New York, USA: PMLR, 2022:4436-4454.
[14]
WORTSMANM, ILHARCOG, KIMJ W, et al. Robust fine-tuning of zero-shot models[C]∥Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2022:7949-7961.
[15]
JIANGZ Y, CHENT L, CHENT, et al. Robust pre-training by adversarial contrastive learning[C]∥Proceedings of the 34th International Conference on Neural Information Processing Systems. Red Hook, USA: Curran Associates Inc., 2020:16199-16210.
[16]
JEDDIA, SHAFIEEM J, WONGA. A simple fine-tuning is all you need: towards robust deep learning via adversarial fine-tuning[DB/OL]. (2020-12-25)[2024-11-05].