Scholarlink

<p>In response to the adversarial vulnerabilities of deep vision algorithms, existing security assessment methods often face problems such as low generation efficiency and severe visual distortion. It is urgent to explore a new paradigm for adversarial sample generation that takes into account high concealment and computational power economy. To this end, this paper aims to overcome the limitations of pixel-level iterative computation capabilities, eliminate biases in biological visual perception, and construct an adversarial example generation model based on latent space distribution search and cross-channel optimization techniques. This method uses a multi-layer convolutional auto-encoder as the base for dimensionality reduction, and maps the image matrix to a low-dimensional feature space for distribution search to extract adversarial priors. In terms of cross-channel quality optimization, by combining the human cone cell spectral sensing threshold and edge contour detection mask, perturbation isolation and precise location across color channels are implemented, and the counter-perturbation is covertly redistributed to visual low-sensitivity areas and high-frequency redundant areas. Experimental evaluation shows that this model maintains excellent visual fidelity while significantly reducing computing power overhead. The PSNR of adversarial samples reaches a maximum of 40.2 decibels. When the perturbation budget is increased to a standardized <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(L_{\infty }\)</EquationSource></InlineEquation>&#xa0;norm constraint of 20/255, the perceptual distance metric remains at a low of 0.125. In actual black-box detection, the model’s migration penetration rate for the large unknown residual visual model reaches 78.5%, resulting in a sharp 88.0% drop in the median target prediction confidence. This paper not only provides a high-throughput automated vulnerability investigation tool for ubiquitous artificial intelligence vision applications, but also establishes a solid theoretical foundation for cross-channel physical isolation for robust boundary reconstruction of future defense systems.</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

Efficient generation of adversarial examples and cross-channel quality optimization based on latent space distribution search

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月24日

查看全文
  • Yaoqin Sun

摘要

In response to the adversarial vulnerabilities of deep vision algorithms, existing security assessment methods often face problems such as low generation efficiency and severe visual distortion. It is urgent to explore a new paradigm for adversarial sample generation that takes into account high concealment and computational power economy. To this end, this paper aims to overcome the limitations of pixel-level iterative computation capabilities, eliminate biases in biological visual perception, and construct an adversarial example generation model based on latent space distribution search and cross-channel optimization techniques. This method uses a multi-layer convolutional auto-encoder as the base for dimensionality reduction, and maps the image matrix to a low-dimensional feature space for distribution search to extract adversarial priors. In terms of cross-channel quality optimization, by combining the human cone cell spectral sensing threshold and edge contour detection mask, perturbation isolation and precise location across color channels are implemented, and the counter-perturbation is covertly redistributed to visual low-sensitivity areas and high-frequency redundant areas. Experimental evaluation shows that this model maintains excellent visual fidelity while significantly reducing computing power overhead. The PSNR of adversarial samples reaches a maximum of 40.2 decibels. When the perturbation budget is increased to a standardized \(L_{\infty }\) norm constraint of 20/255, the perceptual distance metric remains at a low of 0.125. In actual black-box detection, the model’s migration penetration rate for the large unknown residual visual model reaches 78.5%, resulting in a sharp 88.0% drop in the median target prediction confidence. This paper not only provides a high-throughput automated vulnerability investigation tool for ubiquitous artificial intelligence vision applications, but also establishes a solid theoretical foundation for cross-channel physical isolation for robust boundary reconstruction of future defense systems.