Scholarlink

<p>In this paper, the effectiveness of approximate computing in enhancing the robustness of deep neural networks (DNNs) against adversarial attacks is investigated. Furthermore, an approximate multiplier based on complementary approximate adders is proposed for providing maximum robustness against adversarial attacks. The proposed approximate multiplier enables energy-efficient and high-throughput inference suitable for large-scale and real-time DNN workloads on HPC systems. Some of the most common adversarial attack algorithms (<i>i.e.</i>, FGSM, BIM, PGD, C&amp;W, Sparse, and AutoAttack) are applied to the input test data of three pretrained DNN models (<i>i.e.</i>, VGG-16, ResNet-34, and DenseNet-121, trained for classifying the CIFAR-10 and SVHN datasets). The adversarial inputs are then applied to the DNN models implemented using exact or different approximate multipliers. The results indicate that the use of approximate multipliers can improve the accuracy of DNNs in the presence of the considered adversarial attacks. Specifically, employing the proposed approximate multiplier in DNN models yields accuracy improvements of up to 72% (30%) over an exact multiplier (other state-of-the-art approximate multipliers) in the presence of adversarial attacks. Our design also provides better hardware characteristics, including delay, power consumption, and area improvements up to 14% (27%), 61% (49%), and 54 (43%), compared to an exact design (state-of-the-art approximate designs), optimizing the DNN models for higher robustness against adversarial attacks.</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

On the use of approximate computing for improving the robustness of DNNs against adversarial attacks

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Sahand Divsalar,
  • Fatemeh Arezoomand,
  • Shaghayegh Vahdat,
  • Ali Afzali-Kusha,
  • Massoud Pedram

摘要

In this paper, the effectiveness of approximate computing in enhancing the robustness of deep neural networks (DNNs) against adversarial attacks is investigated. Furthermore, an approximate multiplier based on complementary approximate adders is proposed for providing maximum robustness against adversarial attacks. The proposed approximate multiplier enables energy-efficient and high-throughput inference suitable for large-scale and real-time DNN workloads on HPC systems. Some of the most common adversarial attack algorithms (i.e., FGSM, BIM, PGD, C&W, Sparse, and AutoAttack) are applied to the input test data of three pretrained DNN models (i.e., VGG-16, ResNet-34, and DenseNet-121, trained for classifying the CIFAR-10 and SVHN datasets). The adversarial inputs are then applied to the DNN models implemented using exact or different approximate multipliers. The results indicate that the use of approximate multipliers can improve the accuracy of DNNs in the presence of the considered adversarial attacks. Specifically, employing the proposed approximate multiplier in DNN models yields accuracy improvements of up to 72% (30%) over an exact multiplier (other state-of-the-art approximate multipliers) in the presence of adversarial attacks. Our design also provides better hardware characteristics, including delay, power consumption, and area improvements up to 14% (27%), 61% (49%), and 54 (43%), compared to an exact design (state-of-the-art approximate designs), optimizing the DNN models for higher robustness against adversarial attacks.