VPFL: A Verifiable Property Federated Learning Framework Against Invisible Attacks in Distributed IoT

The development of Federated Learning (FL) offers an efficient Machine Learning (ML) approach with privacy protection to solve the data island issue in distributed Internet of Things (IoT). However, existing FL frameworks still suffer from invisible attacks in IoT environments, such as free-rider attacks, backdoor attacks, and model theft. In this paper, we propose a Verifiable Property Federated Learning (VPFL) framework to overcome the above invisible attacks. We present a black-box watermarking task distribution mechanism to prevent free-rider attacks by verifying the property of local models. Our adversarial fine-tuning embedding technique can not only eliminate backdoors in global models, but also simultaneously embed white-box watermarks into model parameters to prevent model theft. Comprehensive experimental evaluations demonstrate that our framework outperforms state-of-the-art schemes in terms of security and feasibility against invisible attacks.