研究人员开发了AEGIS，这是一种新颖的防御机制，旨在对抗文本到图像扩散模型中的视觉同义词攻击（VSA）。与以往专注于明确不安全概念的方法不同，AEGIS动态追踪禁止的语义在生成过程中如何出现。通过识别充当不安全视觉语义瓶颈的特定注意力头，AEGIS应用有针对性的排斥来提高安全性和可用性，而不会压制良性概念。该系统已在SD 1.4和SD 2.1等模型上证明了有效性，显著降低了攻击成功率。 AI
影响 增强了文本到图像模型的安全性，可能减少滥用并提高用户信任度。
排序理由 该集群包含一篇详细介绍AI安全新技术的学术论文。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
Text-to-image diffusion models have achieved high visual fidelity and broad adoption, but remain vulnerable to safety violations when adversaries exploit them to synthesize illicit content. Existing alignment paradigms, from input sanitization to structural feature pruning, are l…
arXiv:2607.06120v1 Announce Type: new Abstract: Text-to-image diffusion models have achieved high visual fidelity and broad adoption, but remain vulnerable to safety violations when adversaries exploit them to synthesize illicit content. Existing alignment paradigms, from input s…
Text-to-image diffusion models have achieved high visual fidelity and broad adoption, but remain vulnerable to safety violations when adversaries exploit them to synthesize illicit content. Existing alignment paradigms, from input sanitization to structural feature pruning, are l…