Blockchain and Federated Learning Empowered Digital Twin for Effective Healthcare

Proper exploitation of artificial intelligence (AI) models is required to achieve precise and efficient healthcare. However, this is an obstacle to the possibility of exchanging medical data, as they are considered instances of personal information and subject to data protection legislation and legal requirements. Moreover, because fully automatic decision-making leveraging AI is undesirable, and in contrast to current legislations, human-in-the- loop should be enforced. Digital twins represent a valuable solution to these issues as they allow continuous retraining with newly collected data and seamlessly integrate human intervention within AI-based solutions. However, to address the impossibility of a central node capable of collecting and processing medical data for training an AI model, a network of interacting and collaborating digital twins should be properly defined without leveraging the exchange of medical data to avoid legal issues. Federated learning (FL) alleviates this issue by exchanging model parameters; however, security is compromised by poisoning attacks. This paper presents the architecture of networked digital twins using FL for disease diagnosis across multiple healthcare providers. Moreover, we deal with data-poisoning protection by leveraging the blockchain. The proposed solutions were experimentally assessed to prove their suitability and effectiveness in addressing the introduced research challenges. We prove that under data-poisoning attacks, the achievable accuracy is close to that for the case with no attacks, with a distance of 10%, and is not affected by sudden drops, as in the case of attacks without any protection. We assessed various FL models and achieved an accuracy of approximately 0.8 within an emulated scenario affected by heterogeneities and the impossibility of sharing data samples.