Verifiable secure aggregation scheme for privacy protection in federated learning networks

Federated learning enables multiple participants to construct a distributed machine learning system coordinated by a server. Most existing solutions assume a semi-honest system, considering each participant to be honest but curious, which does not align with the complex real-world environment. In reality, servers might act maliciously by tampering with or forging aggregation results, which directly threatens the integrity of global models.. To verify the integrity of server aggregation computations while protecting the privacy of clients, this paper introduces a privacy-preserving verifiable secure aggregation scheme for federated learning networks. Initially, we construct a functional reuse private key ring generation algorithm, enabling clients to encrypt and protect their private gradients using the private key ring. Subsequently, leveraging the discrete logarithm difficulty problem, we devise a commitment protocol where clients commit to their encrypted private gradients. Upon receiving the aggregation result from the server, they collaboratively unlock the commitment, thereby verifying the aggregation result. Security analysis demonstrates that our solution effectively ensures privacy protection. We tested the performance using a Raspberry Pi as an edge computing device. Experimental data reveals that, with 100 clients, our scheme demonstrates that the additional costs for proof generation and verification computations are 39.9% and 34.1% of the existing scheme, respectively, highlighting its lightweight nature.