Frequency-guided adaptive privacy masking network

Medical image sharing and collaborative analysis have become essential for advancing diagnostic capabilities across healthcare institutions, yet they introduce substantial privacy risks when sensitive patient information is exposed during transmission or processing. Existing approaches to privacy preservation in medical imaging suffer from fundamental limitations: image-level obfuscation methods inevitably degrade diagnostic utility, cryptographic techniques impose prohibitive computational overhead, and federated learning remains vulnerable to gradient-based inference attacks. This paper proposes a Frequency-guided Adaptive Privacy Masking Network (FAPM-Net) for medical image privacy protection, which jointly exploits spatial and frequency information to selectively mask privacy-sensitive components while preserving diagnostic content. The proposed architecture transforms images into the frequency domain, decomposes frequency components into diagnostic and privacy-sensitive branches, generates an adaptive soft attention mask through a lightweight CNN, and applies frequency filtering before reconstruction. The filtered image is simultaneously evaluated by a diagnostic classifier to ensure utility preservation and by an adversarial privacy module to suppress re-identification. Validation on the ChestX-ray14 dataset demonstrates that the proposed method achieves effective privacy protection against re-identification attacks while maintaining diagnostic accuracy comparable to non-private baselines, with the added benefit of computational efficiency suitable for clinical deployment.