Skip to main content
Back to timeline
arXivSource publication:

Backdoors in medical foundation models survive a model swap: BAPLe++ repairs the prompt and cuts attack success more consistently across four models and six datasets

Synopsis

Medical foundation models are often adapted with prompt learning, which trains a few prompt tokens while the backbone stays frozen; BAPLe++ asks whether a backdoor implanted during that adaptation travels with the prompt to a different medical foundation model, and proposes DePLe, which treats the infected prompt as a fixed anchor and learns a new prompt on a small clean set. Across MedCLIP, BioMedCLIP, PLIP and QuiltNet and six chest X-ray and histopathology datasets, backdoors do transfer across models, but asymmetrically and in a dataset-dependent way; compared with four training-free input transformations, DePLe reduces backdoor accuracy more consistently while preserving, and in some settings improving, clean accuracy.

Source-provided article image: BAPLe++: Cross-Model Transferability and Prompt-Level Defense of Backdoors in Medical Foundation Models
arXiv

Interpretation

A backdoor implanted through prompt learning does not stay on its original model: an infected prompt moved to another medical foundation model can still steer triggered inputs to the attacker-chosen class. Prior backdoor work focused on how the attack is built, without evaluating effectiveness when moving from a source medical foundation model to a target one, even though institutions may deploy models trained on different pretraining corpora. Across MedCLIP, BioMedCLIP, PLIP and QuiltNet and six chest X-ray and histopathology datasets, transfer varies with direction, dataset and source backbone: MedCLIP to BioMedCLIP gives a BAD of 0.1, while the reverse weakens considerably; PLIP and QuiltNet transfer roughly symmetrically, weakest on Kather and strongest on DigestPath.

Repair can happen in prompt space alone, without retraining the frozen backbone. The backdoor lives in the prompt rather than the model weights, so discarding the suspect prompt or retraining the model are both unattractive. DePLe keeps the infected prompt as a fixed anchor and learns a new prompt on a small clean set, using a clean-task loss plus a feature-divergence term that pushes class embeddings away from the region induced by the infected prompt; across twelve model-dataset pairs it reaches the lowest backdoor accuracy in most settings, for example cutting MedCLIP on COVID to 0.051 where no input transformation goes below 0.52.

Perspective

The experiments are limited to four CLIP-style dual-encoder medical foundation models and six chest X-ray and histopathology datasets, under a 32-shot setting with a fixed poisoning rate and a frozen backbone; the conclusions apply to backdoor transfer and prompt repair within that scope.

Transfer is not uniform: it depends on direction, dataset and source backbone, so cross-model risk is real but uneven. DePLe is not best everywhere: on MedCLIP with RSNA18, Pixel Deflection gives a marginally lower backdoor accuracy but nearly halves clean accuracy, and on QuiltNet with DigestPath, JPEG-30 attains both lower backdoor accuracy and higher clean accuracy.