Paper 2026/1376

Secure and Efficient Federated Learning with Adaptive Differential Privacy and Verifiable Homomorphic Aggregation

Abstract

Federated Learning (FL) enables collaborative model training without centralizing raw data, but remains vulnerable to gradient inference attacks, malicious aggregation servers, and communication inefficiencies. Existing cryptographic secure aggregation schemes provide confidentiality and verifiability yet lack formal statistical privacy guarantees, while most differential privacy (DP)-based approaches rely on fixed noise injection, resulting in suboptimal privacy--utility tradeoffs. This paper proposes HEAD-FL, a secure and efficient federated learning framework that integrates adaptive differential privacy with verifiable homomorphic aggregation. The proposed scheme introduces a round-adaptive Gaussian perturbation mechanism analyzed under the Rényi Differential Privacy (RDP) framework, enabling tight cumulative privacy accounting and explicit conversion to $(\varepsilon, \delta)$-DP guarantees. By adopting Federated Averaging (FedAvg) instead of gradient-based aggregation, the framework significantly reduces communication overhead while preserving confidentiality, verifiability, and robustness to client dropouts. Theoretical analysis and experimental evaluation demonstrate that HEAD-FLachieves improved privacy--utility tradeoffs and enhanced communication efficiency compared with fixed-noise and gradient-based secure aggregation methods, making it suitable for deployment in privacy-sensitive and bandwidth-constrained environments.