Fed-TTC: A Pareto-Efficient Hierarchical Federated Learning Framework

Federated learning (FL) has emerged as a transformative paradigm for privacy-preserving collaborative machine learning, yet scaling to millions of clients reveals critical bottlenecks in single-server architectures. Multi-server FL addresses these scalability challenges through distributed aggregation, but existing hierarchical FL (HFL) frameworks suffer from inefficient client-server assignment mechanisms that prioritize stability while neglecting Pareto-efficiency (PE) and strategy-proofness (SP). This work fundamentally reformulates the client-server assignment problem through the lens of market design, drawing an analogy to the school choice problem where clients and servers are agents with preferences and priorities. We introduce a novel framework that applies the Top Trading Cycle (TTC) algorithm to HFL, guaranteeing both PE and SP while achieving minimal instability. Our approach integrates a heuristic-based Shapley value method using the One-Round (OR) approximation to fairly quantify client contributions, enabling servers to form priority rankings that reflect actual utility. To eliminate centralization concerns inherent in traditional HFL architectures, we implement the first blockchain-based deployment of key matching mechanisms, such as TTC, Deferred Acceptance (DA), and Immediate Acceptance with Skip (IAS) as the smart contracts, demonstrating the feasibility of trustless client selection. Preliminary results on CIFAR-10 show that TTC combined with OR-based contribution assessment outperforms baseline assignment strategies including DA, IAS, and the MAAIM frameworks in final model accuracy. This doctoral research further proposes investigations into dynamic TTC mechanisms and their integration with FedProx aggregation in flat multi-server architectures, addressing the statistical heterogeneity challenges of decentralized, non-IID environments while maintaining the theoretical guarantees of SP matching.