关于一般化"跳过连接"的对抗性转让性
IEEE transactions on pattern analysis and machine intelligence
|February 18, 2026
概括
深度学习模型中的跳过连接无意中有助于创建可转移的对抗示例. 一个新的跳过梯度方法 (SGM) 增强了这些跨各种架构和域的攻击.
科学领域:
- 人工智能的人工智能
- 计算机视觉 计算机视觉
- 机器学习安全 机器学习安全
背景情况:
- 跳过连接对于在标准条件下深度学习模型的性能至关重要.
- 然而,它们在对抗场景中的作用仍未得到充分研究.
研究的目的:
- 调查跳过连接对对抗性示例可转移性的影响.
- 提出一种新的方法,即跳过梯度方法 (SGM),以利用这种特性.
主要方法:
- 在对抗性攻击下分析ResNet类架构中的梯度流.
- 开发SGM以偏向向后传播到跳过连接梯度.
- 将SGM扩展到各种架构,如视觉转换器 (ViTs) 和自然语言处理 (NLP) 模型.
主要成果:
- SGM显著提高了各种模型家族 (ResNets,Transformers,LLMs) 中对抗性示例的可转移性.
- 这种方法对集体攻击,有针对性的攻击和具有防御的模型仍然有效.
- 经验证据和理论解释支持SGM的有效性.
结论:
- 跳过连接是一个可用于生成高度可转移的对抗性攻击的漏洞.
- SGM为对抗性攻击提供了一种强有力的技术,突出了架构漏洞.
- 这些发现促使人们对安全模型架构设计进行研究.
更多相关视频
08:43Application of Granger Causality Analysis of the Directed Functional Connection in Alzheimer's Disease and Mild Cognitive Impairment
Published on: August 7, 2017
8.5K
11:09RBDT: A Computerized Task System based in Transposition for the Continuous Analysis of Relational Behavior Dynamics in Humans
Published on: July 17, 2021
3.4K
相关概念视频
Woodward–Hoffmann Selection Rules and Microscopic Reversibility
4.1K
Electrocyclic reactions, cycloadditions, and sigmatropic rearrangements are concerted pericyclic reactions that proceed via a cyclic transition state. These reactions are stereospecific and regioselective. The stereochemistry of the products depends on the symmetry characteristics of the interacting orbitals and the reaction conditions. Accordingly, pericyclic reactions are classified as either symmetry-allowed or symmetry-forbidden. Woodward and Hoffmann presented the selection criteria for...
4.1K
Associative Learning
1.5K
Associative learning is a fundamental concept in behavioral psychology, wherein a connection is established between two stimuli or events, leading to a learned response. This process is critical in understanding how behaviors are acquired and modified. Conditioning, the mechanism through which associations are formed, can be divided into two main types: classical conditioning and operant conditioning, each elucidating different aspects of associative learning.
Classical conditioning, also known...
Classical conditioning, also known...
1.5K
Propagation of Action Potentials
9.9K
The propagation of an action potential refers to the process by which a nerve impulse, or "action potential," travels along a neuron.
Neurons (nerve cells) have a resting membrane potential, with a slightly negative charge inside compared to outside. This is maintained by ion channels, such as sodium (Na+) and potassium (K+) channels, which control the flow of ions. When a stimulus, like a touch or a signal from another neuron, triggers the neuron, sodium channels open, allowing sodium ions to...
Neurons (nerve cells) have a resting membrane potential, with a slightly negative charge inside compared to outside. This is maintained by ion channels, such as sodium (Na+) and potassium (K+) channels, which control the flow of ions. When a stimulus, like a touch or a signal from another neuron, triggers the neuron, sodium channels open, allowing sodium ions to...
9.9K
Transfer Function to State Space
832
State-space representation is a powerful tool for simulating physical systems on digital computers, necessitating the conversion of the transfer function into state-space form. Consider an nth-order linear differential equation with constant coefficients, like those encountered in an RLC circuit. The state variables are selected as the output and its n−1 derivatives. Differentiating these variables and substituting them back into the original equation produces the state equations.
In an RLC...
In an RLC...
832
Convolution Properties I
628
Convolution computations can be simplified by utilizing their inherent properties.
The commutative property reveals that the input and the impulse response of an LTI (Linear Time-Invariant) system can be interchanged without affecting the output:
The commutative property reveals that the input and the impulse response of an LTI (Linear Time-Invariant) system can be interchanged without affecting the output:
628
Reynolds Transport Theorem
2.0K
The Reynolds transport theorem provides a framework to relate the time rate of change of an extensive property within a system to that in a control volume, which is crucial for analyzing fluid dynamics. Extensive properties, such as mass, velocity, acceleration, temperature, and momentum, can be expressed in terms of the mass of a fluid portion. These properties are called extensive because they depend on the system's size, while intensive properties are their corresponding values per unit...
2.0K
