Research on cross-model transfer and defense of prompt-learning backdoors in medical foundation models

1. 22:45
   arXiv: Adversarial Attacks, Backdoors & Poisoning

   BAPLe++: Cross-Model Transferability of Backdoors in Medical Foundation Models and Prompt-Level Defense

   The researchers propose BAPLe++, which studies the cross-model transferability of prompt backdoors among medical foundation models, and propose a prompt-level defense method, DePLe. Medical foundation models are often adapted to clinical tasks through prompt learning, fine-tuning only a small number of prompt tokens while freezing the backbone; prior research showed that in few-shot adaptation, jointly optimizing learnable prompts and imperceptible noise triggers can implant a backdoor, directing triggered inputs to an attacker-chosen class without affecting clean performance. The paper proposes two metrics, CAD and BAD, and the authors claim this is the first study on cross-model transferability of prompt backdoors; experiments on four medical foundation models and six chest X-ray and histopathology datasets show that backdoors can transfer across models, but asymmetrically and depending on the dataset. DePLe uses the infected prompt as a fixed anchor, learns a new prompt on a small clean set, and deviates from the infected embedding region while preserving the clean task; compared with four training-free input transformations, DePLe reduces backdoor accuracy more consistently while preserving and, in several settings, improving clean accuracy.

Current attention 34·Peak in the comparable period 37 (Nov 10 21:00)·Change over the comparable past 24 hours 0%