BioShield: Domain-Adaptive Backdoor Defence for Latent Diffusion Models in Clinical Medical Imaging
Year: 2026
Type: conference-paper
Abstract: Latent diffusion models (LDMs) are increasingly used for synthetic medical image generation, yet their vulnerability to backdoor attacks in multi-domain clinical settings remains underexplored. We present BioShield, an inference-time defence that combines three modules: Adversarial Noise Smoothing (ANS) for suppressing high-frequency trigger artefacts, Trigger Latent Inversion (TLI) for Mahalanobis-based latent anomaly detection and correction, and Spatial Attention Unlearning (SAU) for suppressing trigger-correlated attention patterns in the U-Net denoiser without modifying model weights. A lightweight domain classifier routes inputs to domain-specific prototype banks, enabling consistent operation across chest radiography and retinal fundus imaging. Evaluated against five backdoor attacks on a controlled LDM testbed, BioShield achieves mean TLI detection AUROC of 0.656, FID of 0.08 on purified outputs versus 8.66 for FIBA-triggered images, PSNR of 26.5–27.2 dB, and total inference overhead of 11.14×. Cross-domain misrouting increases the false-positive rate from 5.2% to 100%, showing that domain routing is structurally necessary in multi-domain deployment. We further analyse a trajectory-variance extension (TLI-NSv), which improves detection for selected attacks but yields a net mean AUROC regression of −0.031, indicating that naive ensemble fusion is unstable across heterogeneous trigger types.
Authors: Md Mynoddin, M Abdur Rabbi Tota, Nusrat Jahan
Institution: Rangamati Science and Technology University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204436756