1 paper across 1 session
We detect and remove backdoor samples in MLLM fine-tuning by identifying abnormal attention entropy patterns without requiring clean data or model modifications.
