Scholarlink

Convolutional Neural Networks (CNNs) have achieved remarkable accuracy in image classification tasks, but their susceptibility to adversarial examples remains a significant concern. While gradient-based methods such as FGSM, I-FGSM, and PGD are standard for evaluating robustness, the role of image resolution in modulating attack effectiveness has received limited attention. In this paper, we present a systematic evaluation of adversarial robustness across three neural network architectures (i) standard CNNs, (ii) dilated CNNs, and (iii) Vision Transformers (ViTs)—using both low-resolution (FMNIST) and high-resolution (ANIMAL5) datasets. To better capture the influence of resolution, we train and evaluate models on ANIMAL5 at four granularities: 28 \(\,\times \,\) 28, 64 \(\,\times \,\) 64, 112 \(\,\times \,\) 112, and 224 \(\,\times \,\) 224. In addition, we test adversarial performance across a range of \(\epsilon \) , \(\alpha \) and iteration values, balancing attack strength and computational efficiency. Our results demonstrate that CNN-based models are more vulnerable to attacks at higher resolutions, while ViTs behave in the opposite trend, demonstrating greater susceptibility at lower resolutions. In addition, the results show that image resolution affects architectures differently, even among similar models, depending on the underlying mechanisms they employ, such as dilation or attention.

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

Resolution Matters: Evaluating Adversarial Robustness of CNN and ViT Models Under Varying Image Resolutions

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sachin Sharma,
  • Miltiadis Alamaniotis,
  • Jiho Noh,
  • Michail S Alexiou

摘要

Convolutional Neural Networks (CNNs) have achieved remarkable accuracy in image classification tasks, but their susceptibility to adversarial examples remains a significant concern. While gradient-based methods such as FGSM, I-FGSM, and PGD are standard for evaluating robustness, the role of image resolution in modulating attack effectiveness has received limited attention. In this paper, we present a systematic evaluation of adversarial robustness across three neural network architectures (i) standard CNNs, (ii) dilated CNNs, and (iii) Vision Transformers (ViTs)—using both low-resolution (FMNIST) and high-resolution (ANIMAL5) datasets. To better capture the influence of resolution, we train and evaluate models on ANIMAL5 at four granularities: 28 \(\,\times \,\) 28, 64 \(\,\times \,\) 64, 112 \(\,\times \,\) 112, and 224 \(\,\times \,\) 224. In addition, we test adversarial performance across a range of \(\epsilon \) , \(\alpha \) and iteration values, balancing attack strength and computational efficiency. Our results demonstrate that CNN-based models are more vulnerable to attacks at higher resolutions, while ViTs behave in the opposite trend, demonstrating greater susceptibility at lower resolutions. In addition, the results show that image resolution affects architectures differently, even among similar models, depending on the underlying mechanisms they employ, such as dilation or attention.