Spectral and Activation Clustering Combined for Backdoor Detection in Medical Imaging Models

A new arXiv paper proposes combining spectral signatures of model weights with clustering of internal activations to detect backdoors planted in healthcare imaging models during training. The authors frame the work around sector-level guidance that names model poisoning and adversarial manipulation as concerns for clinical machine learning deployments. The paper covers the method, its implementation, and an evaluation of detection performance.