Federated MRI-to-Synthetic CT Translation Vulnerable to Membership Inference Attacks, Secure Aggregation Mitigates Risk

Membership inference attacks on federated MRI-to-synthetic CT translation achieved AUC 0.92–0.99 across centers. Secure aggregation reduced AUC to near-random levels (0.23–0.56) without impacting synthesis quality.

• Deep leakage from gradients recovered only coarse anatomical structures (SSIM ≤0.16, PSNR ≤11 dB), with no clinically identifiable details.
• Under high-noise data poisoning, standard federated averaging rendered the federation inoperative; robust median aggregation restored near-baseline performance in most scenarios.
• Low-level poisoning could be stealthy, mimicking natural heterogeneity across centers.

Automated summary

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