Lightweight multi-layered de-identification architecture: Secure client selection in federated learning

Despite federated learning (FL) let us avoid raw data-sharing, recent studies shows that existing client selection schemes poses privacy vulnerabilities that might be led to organizational identity leaks. To address this problem, we present a lightweight, multi-layered de-identification architecture that enables privacy-preserving client selection while maintaining selection efficiency in FL environments. The architecture comprises lightweight dynamic hash-based identifier, a secure salt distribution protocol, and an enhanced, resizable Bloom filter verifier. Together, these components provide three critical security properties, formal k-anonymity, $ϵ$-connection resistance, and $\alpha$-membership confusion, protecting FL training against dictionary attacks, cross-round linkability attacks, and membership-inference attacks. We also offer a theoretical framework that allows fine-grained control of the privacy-efficiency trade-off through adjustable parameters. Empirical experiment on MNIST, Fashion-MNIST, CIFAR-10) with Non-IID datasets and heterogeneous client environments show that our method keeps model accuracy within 0.02 to 15.3 pp of FedAvg and Power-of-Choice while adding only 3.6% to 6.3% of computation overhead. These results demonstrate the effectiveness of our approach and show that balanced client selection between performance and privacy is achievable.