Ebtisam Alshehri

Defending Against Harmful Supervision Hidden in Benign Samples

Bang An, Yibo Yang, Dandan Guo, Ebtisam Alshehri, Carlos Hinojosa, Bernard Ghanem

Existing defenses are effective when harmful content is explicitly mixed into downstream fine-tuning data, but crafted samples can instead hide harmful supervision inside benign tasks. We propose Embedded Attack, where harmful QA pairs are embedded within benign training samples,…Read more