安全的幻觉? 在扩散模型中测试概念去除的边界
概括
概念删除方法无法从文本到图像扩散模型中完全删除敏感内容. 一个新的概念嵌入对手 (CEA) 攻击可以再生被删除的概念,如NSFW图像和艺术风格.
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 机器学习
背景情况:
- 文本到图像扩散模型产生高质量的视觉效果,但会带来安全风险.
- 潜在的风险包括生成不安全工作 (NSFW) 内容,风格复制和深度假冒.
- 概念删除技术旨在通过删除敏感概念来减轻这些风险.
研究的目的:
- 评估概念删除方法对抗对抗攻击的稳定性.
- 提出一个新的框架,用于在扩散模型中再生被删除的概念.
主要方法:
- 引入了一个称为概念嵌入对手 (CEA) 的两阶段优化攻击框架.
- 利用CLIP模型中的交叉模式对齐来调整对抗性嵌入矢量.
- 开发了欺骗性的对抗性提示,以迫使扩散模型再生被删除的概念.
主要成果:
- 现有的概念删除方法很容易受到针对已删除概念的对抗性提示.
- CEA成功地再生了被删除的概念,包括NSFW内容,艺术风格和对象.
- 在重建已删除的概念方面,CEA取得了显著更高的对抗性成功率 (ASR).
结论:
- 目前的概念删除技术不足以完全消除敏感的概念.
- 该CEA框架有效地利用扩散模型中的潜在空间漏洞.
- 需要强有力的安全措施来解决当前概念删除方法的局限性.
更多相关视频
06:42Continuous Theta Burst Stimulation of the Posterior Medial Frontal Cortex to Experimentally Reduce Ideological Threat Responses
Published on: September 28, 2018
12.1K
07:34Perceptual and Category Processing of the Uncanny Valley Hypothesis' Dimension of Human Likeness: Some Methodological Issues
Published on: June 3, 2013
17.9K
相关概念视频
Theories of Dissolution: Diffusion Layer Model
1.6K
Dissolution, the process by which drug particles dissolve in a solvent, is explained by the diffusion layer model, a theoretical framework that simulates the absorption of oral drugs and allows us to analyze experimental data.
This process starts with a thin layer, saturated with the drug, forming at the interface between the solid and liquid. The solute then diffuses from this layer into the main solution. The Noyes-Whitney equation suggests that the rate of dissolution relies on the diffusion...
This process starts with a thin layer, saturated with the drug, forming at the interface between the solid and liquid. The solute then diffuses from this layer into the main solution. The Noyes-Whitney equation suggests that the rate of dissolution relies on the diffusion...
1.6K
Theories of Dissolution: The Danckwerts' Model and Interfacial Barrier Model
764
Various dissolution theories provide insight into the factors that influence the dissolution rate. Danckwerts' Model suggests that turbulence, rather than a stagnant layer, characterizes the dissolution medium at the solid-liquid interface. In this model, the agitated solvent contains macroscopic packets that move to the interface via eddy currents, facilitating the absorption and delivery of the drug to the bulk solution. The regular replenishment of solvent packets maintains the...
764
Deindividuation
30.3K
Deindividuation is a form of social influence on an individual’s behavior such that the individual engages in unusual or non-normal behavior while in a group setting. Why? Because in these group settings, the individual no longer sees themselves as an individual anymore, disinhibiting their behavior and personal restraint.
30.3K
Magical Thinking
186
Magical thinking encompasses the belief in assumptions that defy logical reasoning yet appear intuitively convincing. It is a common psychological phenomenon that persists across various cultural and individual contexts. While these assumptions contradict empirical evidence and scientific laws, they often serve meaningful psychological roles in promoting emotional resilience and a sense of control, especially under stress or uncertainty.Thought-Action Fusion and the Law of SimilarityA key...
186
Causes of Similarity-Dissimilarity Effect
253
The similarity-dissimilarity effect, a fundamental concept in social psychology, explains how interpersonal similarities and differences influence attraction and social interactions. This effect is supported by three key psychological perspectives: balance theory, social comparison theory, and consensual validation.Balance Theory and Cognitive ConsistencyBalance theory, developed by Fritz Heider, posits that individuals seek cognitive consistency in their relationships. When two people share...
253
The Scientific Method
64.7K
Research is what makes the difference between facts and opinions. Facts are observable realities, and opinions are personal judgments, conclusions, or attitudes that may or may not be accurate. In the scientific community, facts can be established only using evidence collected through empirical research.
64.7K
