arXiv: HADES: Privacy-Preserving Federated Learning via Selective Feature Encryption and Hybrid Model Fusion

This publication introduces HADES, a novel technical framework for privacy-preserving federated learning that combines selective feature encryption with hybrid model fusion. The paper, released on arXiv in June 2026, proposes a method to protect sensitive data during collaborative machine learning by encrypting only specific data features rather than entire datasets, while still allowing models to be trained across multiple parties. This approach aims to reduce computational overhead and improve accuracy compared to traditional differential privacy or full encryption methods.

The primary impact falls on organizations that handle sensitive personal data and use federated learning, particularly in healthcare, finance, and cross-border AI services operating under EU regulations. Any sector deploying distributed machine learning where data cannot be centralized due to GDPR, the AI Act, or sector-specific rules like MDR or PSD2 should review this technique. It offers a potential path to comply with data minimization and purpose limitation principles while maintaining model performance.

Compliance teams should first assess whether their current federated learning implementations rely on full encryption or differential privacy, as HADES may offer a more efficient alternative. They should engage with data protection officers and technical leads to evaluate if selective feature encryption aligns with their data processing records and risk assessments. Finally, monitor regulatory guidance from the EDPB and national authorities on whether such selective encryption methods meet the standard of appropriate technical measures under Article 32 GDPR, and consider updating data protection impact assessments accordingly.