Abstract: To address the problem where image classification accuracy of ResNet （Residual Network） drops due to unknown heavy-tailed noise, a Multi-distribution Heavy-Tailed Noise Adaptive ResNet (MHTNA-ResNet) model was proposed. Firstly, to mitigate the impact of heavy-tailed noise on final predictions, a Multi-distribution Heavy-Tailed Noise Adaptive Layer (MHTNA) was designed, which created noise templates using various heavy-tailed distributions to perturb clean training data, enabling ResNet to learn recognition capabilities for heavy-tailed noisy images through training. Secondly, MHTNA was adaptively trained up-dating noise templates and using the maximum likelihood estimation method to solve for noise template parameters. Then, noise templates were regenerated based on these parameters to ensure adherence to a heavy-tailed distribution. Finally, during testing, the MHTNA was masked to simulate heavy-tailed noise attacks on test images, thereby evaluating the model's capability of noise resistance. Compared to the PRIME method, the classification accuracy of MHTNA-ResNet improves by an average of 3.94, 7.09 and 5.46 percentage points on the CIFAR-10, CIFAR-100 and Mini-ImageNet datasets when faced with heavy-tailed noise attacks. The experimental results demonstrate that the proposed method can effectively improve ResNet's robustness against heavy-tailed noise interference. Abstract: To address the problem where image classification accuracy of ResNet （Residual Network） drops due to unknown heavy-tailed noise, a Multi-distribution Heavy-Tailed Noise Adaptive ResNet (MHTNA-ResNet) model was proposed. Firstly, to mitigate the impact of heavy-tailed noise on final predictions, a Multi-distribution Heavy-Tailed Noise Adaptive Layer (MHTNA) was designed, which created noise templates using various heavy-tailed distributions to perturb clean training data, enabling ResNet to learn recognition capabilities for heavy-tailed noisy images through training. Secondly, MHTNA was adaptively trained up-dating noise templates and using the maximum likelihood estimation method to solve for noise template parameters. Then, noise templates were regenerated based on these parameters to ensure adherence to a heavy-tailed distribution. Finally, during testing, the MHTNA was masked to simulate heavy-tailed noise attacks on test images, thereby evaluating the model's capability of noise resistance. Compared to the PRIME method, the classification accuracy of MHTNA-ResNet improves by an average of 3.94, 7.09 and 5.46 percentage points on the CIFAR-10, CIFAR-100 and Mini-ImageNet datasets when faced with heavy-tailed noise attacks. The experimental results demonstrate that the proposed method can effectively improve ResNet's robustness against heavy-tailed noise interference. 

Key words: heavy-tailed noise, image classification, Residual Network (ResNet), multi-distribution, deep learning
 

摘要： 针对残差网络(ResNet)在图像分类中容易受未知重尾噪声影响导致识别准确率下降的问题，提出一种多分布重尾噪声自适应残差网络模型(MHTNA-ResNet)。首先，为抑制重尾噪声对最终预测的影响，设计一个多分布重尾噪声自适应层(MHTNA)，该层使用多种重尾分布创建噪声模板，对干净的训练数据进行扰动，使ResNet通过训练获得对重尾噪声图像的识别能力；其次，MHTNA在训练中进行自适应训练，使用最大似然估计法求解更新的噪声模板参数，并根据求解参数重新生成噪声模板，控制噪声始终遵循重尾分布；最后，测试时屏蔽MHTNA，对测试图像进行重尾噪声攻击，从而检验模型的抗噪能力。与PRIME方法相比，面对重尾噪声的攻击，在CIFIA10、CIFIA100和MINI-ImageNet数据集上所提方法的分类准确率分别平均提升了3.94、7.09和5.46个百分点。实验结果表明所提方法可以有效地提高ResNet面对重尾噪声干扰时的鲁棒性。

关键词: 重尾噪声, 图像分类, 残差网络, 多分布, 深度学习