Novel invisible poisoning attack and holistic defense for medical image vision transformers

Novel invisible data-poisoning attack for medical image AI achieved high invisibility (PSNR 43.46 dB, SSIM 0.9925); holistic defense detected poisoned samples with 95.8% accuracy (F1 >0.95) in a simulation across radiology, ophthalmology, and pathology datasets.

• Study proposed an invisible poisoning attack using edge-based triggers injected via a Deep Multi-Scale U-Net, and a defense with an attention-aware GAN mimic model; evaluated on radiology, ophthalmology, and pathology images (sample size not reported).
• Defense maintained detection accuracy up to 95.8% (F1 >0.95) across static poisoning rates of 5–30%, improving up to 17% over baseline methods at high contamination.
• Limited to simulated poisoning scenarios; no external validation or real-world deployment testing was described.

RadPigeon summaries are original and for information only. They are not clinical advice.