Agent-oriented Joint Decision Support for Data Owners in Auction-based Federated Learning

TL;DR

This work addresses the lack of data-owner (DO) decision support in Auction-based Federated Learning (AFL) by introducing PAS-AFL, an agent-oriented framework that uses Lyapunov optimization to jointly decide on FL task acceptance $x_i(t)$, sub-delegation $s_i(t)$, and pricing $p_i(t)$ for each DO. By modeling DO decision dynamics with virtual queues and a Lyapunov drift objective, PAS-AFL achieves stable throughput while maximizing per-slot utility, enabling DOs to take on multiple FL tasks concurrently. The approach leverages each DO's reputation $r_i(t)$, pending workload, trust network $oldsymbol{n}_i$, and willingness to train, leading to a per-slot strategy that decouples into optimal $s_i(t)$, $p_i(t)$, and $x_i(t)$. Empirical results on six benchmarks show substantial gains in DO utility (average +28.77%) and FL model accuracy (average +2.64%), underscoring PAS-AFL's potential to enhance DO participation and AFL ecosystem throughput.

Abstract

Auction-based Federated Learning (AFL) has attracted extensive research interest due to its ability to motivate data owners (DOs) to join FL through economic means. While many existing AFL methods focus on providing decision support to model users (MUs) and the AFL auctioneer, decision support for data owners remains open. To bridge this gap, we propose a first-of-its-kind agent-oriented joint Pricing, Acceptance and Sub-delegation decision support approach for data owners in AFL (PAS-AFL). By considering a DO's current reputation, pending FL tasks, willingness to train FL models, and its trust relationships with other DOs, it provides a systematic approach for a DO to make joint decisions on AFL bid acceptance, task sub-delegation and pricing based on Lyapunov optimization to maximize its utility. It is the first to enable each DO to take on multiple FL tasks simultaneously to earn higher income for DOs and enhance the throughput of FL tasks in the AFL ecosystem. Extensive experiments based on six benchmarking datasets demonstrate significant advantages of PAS-AFL compared to six alternative strategies, beating the best baseline by 28.77% and 2.64% on average in terms of utility and test accuracy of the resulting FL models, respectively.

Figures (2)

• Figure 1: An example trust network among agents. The purple nodes denote the agents who are delegating/sub-delegating FL tasks. The green nodes represent the DO agents. The size of the node corresponds to its connectivity gao2023multi.
• Figure 2: The PAS-AFL system architecture.