New backdoor bypasses AI concept erasure, exposes harmful content

Researchers have identified a significant vulnerability in concept erasure techniques designed for text-to-image diffusion models, termed the Erasure Evasion Backdoor (EEB). This backdoor allows adversaries to embed a hidden trigger linked to a concept slated for removal, ensuring that harmful content associated with that concept can still be generated even after erasure attempts. The EEB was shown to be effective across multiple state-of-the-art erasure methods, leading to substantial success rates in generating unwanted outputs, including celebrity likenesses and explicit imagery. AI

IMPACT Highlights a critical flaw in AI safety mechanisms, necessitating new methods to ensure genuine concept removal and prevent misuse.

RANK_REASON The cluster contains a research paper detailing a new vulnerability in AI model safety techniques. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Erasure Evasion Backdoor
• text-to-image diffusion models
• Tobias Braun

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →