Abstract
Federated learning (FL), which addresses data privacy issues by training models on resource-constrained mobile devices in a distributed manner, has attracted significant research attention. However, the problem of optimizing FL client selection in mobile federated learning networks (MFLNs), where devices move in and out of each others' coverage and no FL server knows all the data owners, remains open. To bridge this gap, we propose a first-of-its-kind Socially-aware Federated Client Selection (SocFedCS) approach to minimize costs and train high-quality FL models. SocFedCS enriches the candidate FL client pool by enabling data owners to propagate FL task information through their local networks of trust, even as devices are moving into and out of each others' coverage. Based on Lyapunov optimization, we first transform this time-coupled problem into a step-by-step optimization problem. Then, we design a method based on alternating minimization and self-adaptive global best harmony search to solve this mixed-integer optimization problem. Extensive experiments comparing SocFedCS against five state-of-the-art approaches based on four real-world multimedia datasets demonstrate that it achieves 2.06% higher test accuracy and 12.24% lower cost on average than the best-performing baseline.
Introduction
Federated Learning (FL) has emerged as a transformative paradigm for training artificial intelligence models on local devices, such as smartphones and IoT sensors, without compromising user privacy by centralizing raw data. However, the transition from theoretical FL to practical deployment in Mobile Federated Learning Networks (MFLNs) introduces significant challenges. Unlike static environments, MFLNs are characterized by high mobility and limited server visibility; a central server often only knows a fraction of potential data owners, and these devices frequently move in and out of communication range.
Selecting the optimal subset of clients is critical to model performance and energy efficiency. Existing selection methods often assume a "perfect world" where the server has a global view of all available clients. In reality, the most valuable data for a specific task might reside on a device unknown to the server. To address this, we introduce SocFedCS (Socially-aware Federated Client Selection), a framework that leverages social trust to discover "hidden" data owners and optimize client selection in dynamic mobile environments.
The SocFedCS Framework
Multi-Tier Social Discovery
SocFedCS moves beyond the traditional single-tier client selection model by introducing a two-tier discovery mechanism:
- First-Order Clients (FCs): Devices directly known to and reachable by the FL server.
- Second-Order Clients (SCs): Devices known only to the FCs through local social connections.
By utilizing "trust transitivity," the server can reach high-quality SCs via recommendations from FCs. This effectively expands the candidate pool, allowing the system to access a more diverse and representative dataset than would be possible through direct discovery alone.
Social Trust and Community Impact
To ensure that recommendations are reliable, SocFedCS quantifies social relationships using a Community Impact Factor. This metric is derived from communication frequency and historical interaction data. By prioritizing clients with strong social ties, the system filters for reliability, mitigating the risks associated with "stragglers" or low-quality data updates.
Optimization Strategy
Optimizing client selection in MFLNs is a complex, time-coupled problem—decisions made in one round affect the availability and energy states of devices in future rounds. SocFedCS addresses this using Lyapunov optimization, a mathematical framework that decomposes long-term objectives (such as maintaining a minimum participation rate) into a series of independent, step-by-step optimization problems.
The resulting challenge is a mixed-integer optimization problem: the system must decide whether to select a client (a binary decision) while simultaneously optimizing local model accuracy. We solve this using a hybrid approach:
- Alternating Minimization: To handle the interdependencies between selection and accuracy variables.
- Self-adaptive Global Best Harmony Search (SGHS): A meta-heuristic algorithm used to fine-tune the selection process and ensure the global model converges efficiently.
Experimental Evaluation
Setup
We evaluated SocFedCS using four standard computer vision benchmarks: MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100. To simulate a realistic mobile environment, we employed the Gauss-Markov mobility model to simulate device movement and the Erdos-Renyi random graph to establish initial social ties between users.
Performance Metrics
SocFedCS was compared against five state-of-the-art baselines: Random, Greedy, PowCS, FedCS, and Oort. The results demonstrated superior performance across all datasets:
- Operational Cost: SocFedCS achieved a 12.24% reduction in time-average costs compared to the Greedy baseline, the next best performer.
- Model Accuracy: The framework achieved an average 2.06% higher test accuracy than FedCS.
- Robustness: The system maintained high performance in both IID (Independent and Identically Distributed) and non-IID scenarios, the latter of which is a common failure point for traditional FL algorithms.
Discussion and Insights
The success of SocFedCS highlights the power of the "hidden" client pool. By enabling a recommendation tier, the system finds a "sweet spot" between aggressive selection (which drains battery and increases cost) and conservative selection (which limits accuracy). Furthermore, the use of social trust acts as a natural quality filter, ensuring that the expanded pool contributes positively to the global model.
Conclusion and Future Work
SocFedCS represents a significant advancement in making Federated Learning viable for chaotic, real-world mobile networks. By leveraging social discovery and robust mathematical optimization, it achieves higher accuracy at a lower operational cost.
Future research will focus on extending this framework to support longer chains of trust transitivity (multi-hop recommendations) and investigating the impact of variable recommendation costs ($C_0$) influenced by network congestion and social dynamics. These extensions will further enhance the scalability and depth of client discovery in massive-scale mobile networks.
Summary updated