Mitigating Bias in Client Selection for Federated Learning Using Verifiable Random Functions

Federated Learning (FL) poses unique challenges related to the trustworthiness of both client data and the aggregator. While Secure Aggregation (SA) reduces privacy risks by masking local model updates through cryptographic techniques, it remains susceptible to biased selection attacks. These attacks occur when a malicious aggregator manipulates the client selection process to infer individual model updates, exploiting the lack of verifiability in existing random selection mechanisms. Verifiable Random Functions (VRFs) present a promising solution by ensuring both randomness and verifiability. However, current implementations often rely on blockchain-based random number generation, which incurs significant latency and energy consumption, rendering them impractical for resource-constrained FL environments. In this preliminary research, we propose a lightweight, VRF-based client selection protocol that eliminates the overhead associated with blockchain. Our approach involves a simple interaction between the aggregator and clients to exchange compact random numbers, which are then used to generate unique random tokens for each FL round. We demonstrate that the proposed method achieves statistically unbiased and fully verifiable client selection, even in adversarial scenarios where the aggregator is compromised or colludes with clients.