Improving participants’ competitiveness in an open federated learning environment

Le résumé fourni par la source

Federated learning (FL) is a novel distributed machine learning approach that safeguards data owners’ privacy by enabling training directly on their devices using their personal data. In today’s society, privacy concerns have become increasingly prevalent, and FL is seen as a crucial tool in addressing these issues. This is reflected in recent regulations like the Personal Data Protection Act in Singapore and the General Data Protection Regulation in the European Union. However, these benefits of FL do not come without drawbacks. he decentralized nature of FL grants participants significant autonomy in the information they transmit to the master node during model aggregation, introducing potential vulnerabilities exploitable by malicious actors. Furthermore, FL model performance can suffer from inadequate support from data owners with quality data and sufficient computational/communication resources. Therefore, developing robust client selection frameworks and incentive mechanisms is crucial for enhancing performance and mitigating vulnerabilities in open FL environments. Additionally, the framework should be flexible enough to offer opportunities for participants, who previously performed suboptimally, to prove their capability. This is so that the FL system can benefit from a larger pool of data and resources contributed. Incentive mechanisms are critical in FL systems, as they motivate participation and reliable contributions from quality data owners, especially given the significant computational and communication costs incurred by these participants. In this thesis, I will briefly introduce the current research work in FL, as well as the problems and gaps in this existing work. Thereafter, I will present three key studies that aim to close these gaps. The first study includes a client selection method based on stochastic integer programming, which allows for the federation or task owners to hedge against the selection of low or medium-reputation data owners. This method brings two key advantages. Firstly, it allows federations to hire data owners in a cost-effective manner while taking the necessary precautions against the uncertainty of hiring low- or medium-reputation clients. Secondly, it alleviates the problem of herding, as it provides data owners with another opportunity to improve their reputation value. In our simulations, the proposed approach achieved up to 64.26% cost savings compared to other baseline methods. In the second proposed work, an opportunistic incentive mechanism utilizing Lyapunov optimization is introduced to help federations optimize their cost structure even in an open, non-monopolistic, auction-based FL setting. It accounts for the federations’ level of urgency in completing tasks when recruiting data owners. Such an incentive mechanism is critical in maintaining a good relationship with data owners such that they will continue to contribute reliably. To the best of our knowledge, this work was the first decision support approach designed to help federations optimize budget usage in competitive and open auction-based FL markets. It achieves the best trade-off among fairness, cost-effectiveness, and utility, delivering 1.80% higher utility and 62.85% lower cost compared to the best-performing baseline, while maintaining a comparably high level of fairness. In the final work presented, the approach would challenge the assumption that all data owners must be selected from the outset. Instead, the FL task owner can assess the current circumstances and choose the data owners as and when required. Existing approaches and prior research works typically employ a static threshold to determine the eligibility of data owners. In contrast, this study proposes a more flexible framework that permits data owners with lower reputational score to participate in FL training, thereby facilitating an opportunity for reputation improvement. This strategy has led to significant cost savings for the task owner. In our simulated experiments, we show that this approach reduces costs by up to 33.65% and improves total utility by 2.91%. Overall, our experimental results consistently highlights the superiority of the proposed approaches in terms of utility yield and cost savings when compared to existing methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.