New backdoor attack exploits multimodal AI models with natural language triggers

Researchers have developed a new type of backdoor attack called Text-Guided Backdoor (TGB) that targets multimodal pretrained models. Unlike previous attacks that require specific trigger conditions, TGB utilizes naturally occurring words as triggers, making it more stealthy and practical for real-world scenarios. The attack's strength can be adjusted by introducing visual adversarial perturbations, allowing for flexible control over its effectiveness without altering the poisoned data. Experiments on tasks like Composed Image Retrieval and Visual Question Answering demonstrate TGB's ability to exploit security vulnerabilities in these models. AI

IMPACT This research highlights critical security vulnerabilities in multimodal AI models, potentially impacting their safe deployment in real-world applications.

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

Read on arXiv cs.LG →

• Chaojian Yu
• Composed Image Retrieval (CIR)
• Text-Guided Backdoor (TGB)
• visual question answering (VQA)

• paper
• safety

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