Scholarlink

<p>Deep learning models are highly vulnerable to adversarial attacks, which can significantly compromise their reliability, particularly in critical applications. In this work, we analyse the impact of entropy regularisation on adversarial robustness, revealing its effectiveness in enhancing model resilience. Additionally, we propose a novel defence that achieves a strong balance between clean accuracy and adversarial robustness. The defence operates in the feature space rather than the input space. Working in feature space is faster as it has fewer dimensions than the input space. Specifically, we extract features from the final convolutional layers of the selected classifier and process them using an autoencoder trained to remove adversarial noise while preserving essential information. The purified features are then classified using a fully connected network. This feature-based defence significantly reduces computational overhead while improving robustness against adversarial perturbations. We evaluate our approach on the CIFAR-10 dataset against state-of-the-art adversarial attack techniques like PGD and C&amp;W. The results demonstrate that our approach mitigates adversarial attacks while maintaining high classification accuracy on clean samples. Our method achieves 67.50% (PGD-20), 57.15% (PGD-40), and 74.42% (C&amp;W) robust accuracy on the CIFAR-10 dataset, outperforming standard adversarial training and feature scattering while maintaining 91.82% clean accuracy. Unlike many defences that trade accuracy for robustness, our approach remains effective against both PGD and C&amp;W attacks, surpassing existing methods in overall performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Balancing Accuracy and Robustness Through Feature Reconstruction and Entropy Regularisation: Independent Approaches, Unified Insights

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Siddheshwar Kumar,
  • Shashank Srivastava,
  • Shashwati Banerjea

摘要

Deep learning models are highly vulnerable to adversarial attacks, which can significantly compromise their reliability, particularly in critical applications. In this work, we analyse the impact of entropy regularisation on adversarial robustness, revealing its effectiveness in enhancing model resilience. Additionally, we propose a novel defence that achieves a strong balance between clean accuracy and adversarial robustness. The defence operates in the feature space rather than the input space. Working in feature space is faster as it has fewer dimensions than the input space. Specifically, we extract features from the final convolutional layers of the selected classifier and process them using an autoencoder trained to remove adversarial noise while preserving essential information. The purified features are then classified using a fully connected network. This feature-based defence significantly reduces computational overhead while improving robustness against adversarial perturbations. We evaluate our approach on the CIFAR-10 dataset against state-of-the-art adversarial attack techniques like PGD and C&W. The results demonstrate that our approach mitigates adversarial attacks while maintaining high classification accuracy on clean samples. Our method achieves 67.50% (PGD-20), 57.15% (PGD-40), and 74.42% (C&W) robust accuracy on the CIFAR-10 dataset, outperforming standard adversarial training and feature scattering while maintaining 91.82% clean accuracy. Unlike many defences that trade accuracy for robustness, our approach remains effective against both PGD and C&W attacks, surpassing existing methods in overall performance.