SECURITY RISKS IN MODEL MERGING

Summary from the TLDR AI newsletter

As Model Merging (MM) becomes a popular method for combining fine-tuned models without additional training, new security risks emerge. This paper introduces BadMerging, the first backdoor attack specifically targeting MM.