Shadow-Activated Backdoor Attacks on Multimodal Large Language Models

Keyphrases

• Backdoor Attack 100%
• Multimodal Large Language Models 100%
• Security Threats 16%
• Modeller 16%
• Source Code 16%
• External Trigger 16%
• Interaction Process 16%
• Potential Risk 16%
• Attack Scenario 16%
• Desired Behavior 16%
• Backdoor 16%
• Multi-round 16%
• Malicious Content 16%
• Semantic Discontinuity 16%
• Poisoned Data 16%
• GPT-4 with Vision (GPT-4V) 16%
• Tuning Strategies 16%

Computer Science

• Backdoor Attack 100%
• Multimodal Large Language Models 100%
• Experimental Result 16%
• Interaction Process 16%
• Attack Scenario 16%
• Desired Behavior 16%
• Backdoors 16%
• Malicious Content 16%
• Generative Pre-Trained Transformer 4 16%
• Source Coding 16%