An online adversarial distillation method, based on multi-layer knowledge mutual learning, is proposed to address the issues of limited improvement in robust accuracy, which are caused by the declining reliability of static teacher knowledge and the underutilization of knowledge in existing adversarial distillation methods. A parameter fusion model (PFM) is constructed by weightedly fusing the model parameters of a group of students, with the adversarial loss of PFM being minimized to assist students in finding flatter loss minima. To ensure mutual learning of reliable soft label knowledge at the output layer, students are encouraged to mimic the average clean predictions of all models, in terms of inter-class ranking relationships within the predictive distribution of adversarial examples. Subsequently, to utilize the richer structured knowledge of the middle layer for supplementing the limited information in soft labels, the similarity of sample relationships at the feature channel level among students is maximized. Experimental results demonstrate that, compared to mutual adversarial training (MAT) ResNet-18 network trained by the method exhibits an improvement in test accuracy against adversarial examples generated by projected gradient descent (PGD) by 2.05 and 2.19 percentage point, respectively.
LIAOF Z, LIANGM, DONGY P,et al. Defense against adversarial attacks using high-level representation guided denoiser[C]∥Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2018:1778-1787.
[7]
YANGP, CHENJ, HSIEHC,et al. ML-LOO:detecting adversarial examples with feature attribution[C]∥Proceedings of the 2019 AAAI Conference on Artificial Intelligence. Washington, USA: AAAI, 2019:6639-6647.
[8]
MADRYA, MAKELOVA, SCHMIDTL,et al. Towards deep learning models resistant to adversarial attacks[EB/OL]. [2024-03-04].
[9]
LIUC, SALZMANNM, LINT,et al. On the loss landscape of adversarial training:identifying challenges and how to overcome them[EB/OL]. [2024-03-04].
[10]
GOLDBLUMM, FOWLL H, FEIZIS,et al.Adversarially robust distillation[C]∥Proceedings of the AAAI Conference on Artificial Intelligence. Washington,USA:AAAI,2020,34(4):3996-4003.
[11]
ZIB J, ZHAOS H, MAX J,et al. Revisiting adversarial robustness distillation: robust soft labels make student better[C]∥Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway, USA: IEEE,2021:16423-16432.
[12]
ZHUJ N, YAOJ C, HANB,et al. Reliable adversarial distillation with unreliable teachers[C]∥Proceedings of the 10th International Conference on Learning Representations. Appletom, USA: ICLR, 2022. DOI:10.48550/arXiv.2016.04928 .
[13]
AWAISM, ZHOUF W, XIEC L,et al. MixACM: mixup-based robustness transfer via distillation of activated channel maps[C]∥Proceedings of the Neural Information Processing Systems. Red Hook,USA:Curran Associates Inc.,2021:4555-4569.
[14]
LIUJ, LAUC P, SOURIH, et al. Mutual adversarial training: learning together is better than going alone[EB/OL]. (2021-12-09)[2024-09-01].
[15]
CHEND F, MEIJ P, WANGC,et al. Online knowledge distillation with diverse peers[C]∥Proceedings of the AAAI Conference on Artificial Intelligence. Washington,USA: AAAI, 2020,34(4):3430-3437.
[16]
GOODFELLOWI J, SHLENSJ, SZEGEDYC. Explaining and harnessing adversarial examples[EB/OL]. [2024-03-04].
[17]
CARLININ, WAGNERD.Towards evaluating the robustness of neural networks[C]∥Proceeding of the 2017 IEEE Symposium on Security and Privacy. Piscataway, USA:IEEE, 2017:39-57.
[18]
CROCEF, HEINM. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks[C]∥Proceedings of the International Conference on Machine Learning. New York,USA:ACM,2020:2206-2216.
[19]
ZHANGH Y, YUY D, JIAOJ T,et al. Theoretically principled trade-off between robustness and accuracy[C]∥Proceedings of the International Conference on Machine Learning. New York, USA: ACM, 2019:No.08573.
[20]
ZHANGT L, XUEM Q, ZHANGJ T,et al. Generalization matters: loss minima flattening via parameter hybridization for efficient online knowledge distillation[C]∥Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway,USA:IEEE,2023:20176-20185.
[21]
BLEID M, AN G, JORDANM I. Latent dirichlet allocation[J]. Journal of Machine Learning Research,2003,3:993-1022.
[22]
HUANGT, YOUS, WANGF,et al. Knowledge distillation from a stronger teacher[C]∥Proceedings of the 36th Conference on Neural Information Processing Systems. New York, USA: ACM, 2022:33716-33727.
[23]
BAIY, ZENGY Y, JIANGY, et al. Improving adversarial robustness via channel-wise activation suppressing[C]∥Proceedings of the 9th International Conference on Learning Representations. Appleton, USA: ICLR, 2021. DOI:10.48550/arXiv.2023.08307 .
[24]
GOUJ P, XIONGX S, YUB S,et al. Channel-correlation-based selective knowledge distillation[J]. IEEE Transactions on Cognitive and Developmental Systems,2023,15(3):1574-1585.
[25]
KRIZHEVSKYA, HINTONG. Learning multiple layers of features from tiny images[J]. Handbook of Systemic Autoimmune Diseases, 2009.1(4).DOI: 10.1.1.222.9220 .
[26]
HEK M, ZHANGX Y, RENS Q,et al. Deep residual learning for image recognition[C]∥Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2016:770-778.
[27]
SANDLERM, HOWARDA, ZHUM L,et al. MobileNetV2: inverted residuals and linear bottlenecks[C]∥Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA:IEEE, 2018:4510-4520.