Scholarlink

<p>Deep learning has revolutionized image classification and other AI-driven tasks. Yet, its vulnerability to adversarial attacks remains a critical limitation, particularly in safety–critical domains such as autonomous systems and healthcare. This study proposes the Resilient adversarial network (RANet), a novel defence framework designed to enhance the robustness of deep learning models against adversarial perturbations. RANet integrates adversarial resilience training (ART) through key components: a dedicated adversarial training layer, adaptive perturbation control, feature-space augmentation, and adversarial dropout. These modules collectively improve generalization and reduce susceptibility to both targeted and non-targeted attacks. Evaluated on datasets from the NIPS 2017 adversarial learning challenge, RANet achieves 92.5% accuracy on clean data and 75.3% on adversarial data, outperforming existing defence methods in both accuracy and robustness. The framework demonstrates a strong balance between performance and security, making it viable for real-world deployment.</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

Enhancing Deep Learning with Resilient Adversarial Network (RANet): An Advanced Adversarial Resilience Training Framework for Robust Image Classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月20日

查看全文
  • Saleh Alyahyan

摘要

Deep learning has revolutionized image classification and other AI-driven tasks. Yet, its vulnerability to adversarial attacks remains a critical limitation, particularly in safety–critical domains such as autonomous systems and healthcare. This study proposes the Resilient adversarial network (RANet), a novel defence framework designed to enhance the robustness of deep learning models against adversarial perturbations. RANet integrates adversarial resilience training (ART) through key components: a dedicated adversarial training layer, adaptive perturbation control, feature-space augmentation, and adversarial dropout. These modules collectively improve generalization and reduce susceptibility to both targeted and non-targeted attacks. Evaluated on datasets from the NIPS 2017 adversarial learning challenge, RANet achieves 92.5% accuracy on clean data and 75.3% on adversarial data, outperforming existing defence methods in both accuracy and robustness. The framework demonstrates a strong balance between performance and security, making it viable for real-world deployment.