Secure Service-Oriented Contract Based Incentive Mechanism Design in Federated Learning via Deep Reinforcement Learning

Abstract

In the evolving landscape of federated learning (FL), ensuring the active participation of local model owners (LMOs) while safeguarding data privacy and service security presents a formidable challenge. Our investigation focuses on two different information scenarios: the weakly incomplete information scenario and the strongly incomplete information scenario, which pose unique challenges to the integrity and efficiency of FL systems. In the weakly incomplete information scenario, LMOs have motivations to hide their true types. We use contract theory and exploit its self-revealing properties to ensure LMOs truthfully report their types. In the strongly incomplete information scenario, We present the Contract-based Deep Reinforcement Learning (CDRL) algorithm, which combines the strategic framework of contract theory with the adaptive capabilities of DRL. The CDRL algorithm is designed to perform real-time contract design in dynamic environments, enabling the system to respond effectively to FL participation and ensure continuous alignment of incentives with system security and learning objectives. Through extensive experimentation on a real-world dataset, our proposed mechanism has demonstrated superiority in motivating LMOs to actively participate in FL, thereby significantly improving system performance.