一个空间分布的扰乱策略与光滑的梯度标志方法对图像分类系统的对抗性分析
Yanwei Xu1, Jun Li1, Dajun Chang2
1School of Management Science and Information Engineering, Jilin University of Finance and Economics, Changchun 130117, China.
Entropy (Basel, Switzerland)
|February 27, 2026
概括
本研究介绍了用于深度神经网络 (DNN) 的新对抗性攻击策略 (SD-SGSM). 它通过将干扰聚焦在敏感区域,提高AI系统安全性来增强强性评估.
科学领域:
- 人工智能的人工智能
- 计算机视觉 计算机视觉
- 网络安全 网络安全
背景情况:
- 深度神经网络 (DNN) 在社会技术系统中至关重要,但易受对抗性干扰的影响.
- 当前的攻击方法经常应用统一的扰动,忽视空间灵敏度变化.
- 确保安全关键应用程序中的DNN可靠性,需要对复杂的攻击进行强有力的评估.
研究的目的:
- 开发一个对抗性攻击框架,空间分布式扰动策略与光滑梯度信号方法 (SD-SGSM),以评估DNN的稳定性.
- 通过利用决策依赖区域,最大限度地提高攻击效率,同时最大限度地减少视觉扭曲.
- 通过先进的对抗性测试来提高AI支持系统的安全性和可靠性.
主要方法:
- 拟议的SD-SGSM框架整合了决策依赖域识别,空间适应性扰动分配和渐变平滑.
- 用一个决定性的零出运算符用于关键特征本地化.
- 在微粒度梯度更新中采用过度触角变换.
主要成果:
- 在CIFAR-10数据集上实现了近乎完美的攻击成功率 (99.9%).
- 显著减少了L2扭曲,并保持了高的结构相似性指数指标 (SSIM 0.947).
- 在有效性和视觉准确性方面,超越了现有的单步和基于势头的代攻击.
结论:
- 空间意识,决策依赖的对抗策略对于系统级稳定性评估至关重要.
- 与以前的方法相比,SD-SGSM在攻击强度和视觉质量方面都表现出卓越的性能.
- 调查结果强调,考虑到先进的对抗性威胁,需要安全的AI设计.
相关概念视频
01:19Gradient Vectors and Their Applications
Every point on a topographical map corresponds to a particular elevation, so the landscape can be modeled as a surface whose height depends on horizontal position. From any given location, a hiker may face infinitely many directions, but only one direction produces the fastest possible increase in elevation. This unique route is called the direction of steepest ascent, and in multivariable calculus, it is represented by the gradient vector of the elevation function.The gradient vector points...
01:27Significance of the Gradient Vector
A surface defined by a function of two variables can be understood by examining how it changes along specific directions. When one variable is held constant, the surface reduces to a curve that reflects variation in the other variable. For example, fixing one variable and moving parallel to a coordinate axis produces a cross-sectional curve. The slope of this curve at a given point represents how the function changes in that particular direction, providing a measure of local steepness.By...