New architectural backdoor vulnerability found in Vision-Language Models

Researchers have identified a new type of security vulnerability in Vision-Language Models (VLMs) that can be embedded within the model's architecture. This "architectural backdoor" is inserted through representation steering, allowing malicious actors to subtly alter the model's behavior when a specific trigger is activated, without affecting its performance on clean inputs. The attack can compromise integrity, safety, and fairness across various VLM applications, including question answering and image generation. A proposed auditing defense inspects the model's executable logic rather than just its learned weights. AI

IMPACT Introduces a new class of security threats to AI supply chains, potentially impacting the integrity and safety of deployed models.

RANK_REASON Academic paper detailing a novel security vulnerability in AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• Antonio Emanuele Cinà
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Representation steering
• ScienceCast
• Vision--Language Models

• paper
• safety

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