Scholarlink

<p>The vulnerability of Deep Neural Networks (DNNs) to adversarial attacks reveals critical challenges in model robustness, particularly in sensitive applications. This paper introduces the Lemurs Optimizer for Adversarial Attacks (LO-Attack), a gradient-free metaheuristic algorithm inspired by swarm intelligence and lemur social behavior designed to generate adversarial examples. Unlike traditional gradient-based methods, LO-Attack operates effectively in white-box and black-box environments. Through evaluations on CIFAR-10, ImageNet, and MNIST, LO-Attack achieves a high fooling rate with minimal perturbation sizes, achieving competitive fooling rates compared to FGSM and DeepFool. Visualizations confirm that LO-Attack produces subtle perturbations that compromise model reliability without perceptible differences. The LO-Attack method offers a flexible and robust approach for adversarial attacks and has broader implications for NP-hard optimization challenges beyond neural network testing.</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

Harnessing the Lemurs optimizer for adversarial attacks on deep neural networks

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Ammar Kamal Abasi,
  • Moayad Aloqaily,
  • Mohsen Guizani

摘要

The vulnerability of Deep Neural Networks (DNNs) to adversarial attacks reveals critical challenges in model robustness, particularly in sensitive applications. This paper introduces the Lemurs Optimizer for Adversarial Attacks (LO-Attack), a gradient-free metaheuristic algorithm inspired by swarm intelligence and lemur social behavior designed to generate adversarial examples. Unlike traditional gradient-based methods, LO-Attack operates effectively in white-box and black-box environments. Through evaluations on CIFAR-10, ImageNet, and MNIST, LO-Attack achieves a high fooling rate with minimal perturbation sizes, achieving competitive fooling rates compared to FGSM and DeepFool. Visualizations confirm that LO-Attack produces subtle perturbations that compromise model reliability without perceptible differences. The LO-Attack method offers a flexible and robust approach for adversarial attacks and has broader implications for NP-hard optimization challenges beyond neural network testing.