Adaptive Model Partitioning, Encryption Configuration, and Resource Allocation for Device-and-Edge Collaborative Homomorphic Encrypted Split Federated Learning

Abstract

The integration of homomorphic encryption (HE) into split federated learning (SFL) offers significant security advantages by maintaining end-to-end encryption of intermediate computations. Existing SFL approaches often overlook encryption, adopting fixed machine learning (ML) model partitioning strategies without considering different encryption demands of the users and joint computing and communication resource management of the networks. To this end, this paper proposes an adaptive homomorphic encrypted SFL (HESFL) scheme in device and edge collaborative networks. A training latency minimization problem is formulated to jointly optimize model partitioning, HE configuration, and computing and communication resource allocation. We design a hybrid optimization algorithm combining the proximal policy optimization (PPO) and the sequential least squares programming (SLSQP), where SLSQP derives optimal resource allocation while PPO learns near-optimal partitioning and encryption policies. Experiments are conducted under representative MNIST-based settings, where the HE performance is measured using OpenFHE and TenSEAL, the computing and communication overheads are evaluated under varying security levels. The results demonstrate the superiority of our scheme over three comparative schemes across four scenarios.