New attack injects persistent bias into multimodal AI models

Researchers have developed a novel attack called the Persistent Fairness Backdoor Attack (PFBA) designed to inject and maintain group-specific discrimination into Multimodal Large Language Models (MLLMs). This attack addresses the challenge that standard backdoors degrade with continual learning updates. PFBA utilizes Latent Space Fairness Reinforcement to manipulate the model's feature geometry, preserving utility while amplifying discrimination, and employs Continual Learning Simulation to ensure the backdoor's persistence through future updates. Experiments show that PFBA successfully induces severe and persistent fairness disparities that evade common backdoor defenses. AI

IMPACT Highlights a new vulnerability in MLLMs, potentially impacting their safe deployment in sensitive applications and requiring new defense mechanisms.

RANK_REASON The cluster contains an academic paper detailing a new attack method against AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Multimodal Large Language Models
• Persistent Fairness Backdoor Attack

• safety
• paper
• policy

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