Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection

1. 1Perform causal mediation analysis to identify neurons and layers responsible for unsafe behaviors
2. 2Apply dual-modal safety subspace projection to repair unsafe channels
3. 3Evaluate the safety and performance of the repaired model
4. 4Refine the framework based on experimental results

AI engineers and researchers working on Vision-Language Models can benefit from this framework to improve model safety and reliability, while data scientists can apply the causal mediation analysis to identify unsafe behaviors

💡 Causal discovery and dual-modal safety subspace projection can be used to identify and repair unsafe channels in Vision-Language Models

🚨 Diagnose and repair unsafe channels in Vision-Language Models with CARE framework 💡