Paper 2024/476

OPSA: Efficient and Verifiable One-Pass Secure Aggregation with TEE for Federated Learning

Abstract

In federated learning, secure aggregation (SA) protocols like Flamingo (S\&P'23) and LERNA (ASIACRYPT'23) have achieved efficient multi-round SA in the malicious model. However, each round of their aggregation requires at least three client-server round-trip communications and lacks support for aggregation result verification. Verifiable SA schemes, such as VerSA (TDSC'21) and Eltaras et al.(TIFS'23), provide verifiable aggregation results under the security assumption that the server does not collude with any user. Nonetheless, these schemes incur high communication costs and lack support for efficient multi-round aggregation. Executing SA entirely within Trusted Execution Environment (TEE), as desined in SEAR (TDSC'22), guarantees both privacy and verifiable aggregation. However, the limited physical memory within TEE poses a significant computational bottleneck, particularly when aggregating large models or handling numerous clients. In this work, we introduce OPSA, a multi-round one-pass secure aggregation framework based on TEE to achieve efficient communication, streamlined computation and verifiable aggregation all at once. OPSA employs a new strategy of revealing shared keys in TEE and instantiates two types of masking schemes. Furthermore, a result verification module is designed to be compatible with any type of SA protocol instantiated under the OPSA framework with weaker security assumptions. Compared with the state-of-the-art schemes, OPSA achieves a 2$\sim$10$\times$ speedup in multi-round aggregation while also supporting result verification simultaneously. OPSA is more friendly to scenarios with high network latency and large-scale model aggregation.