TrustFLow: A traceable federated learning framework to enable trustworthy digital twins

As Digital Twin (DT) ecosystems increasingly rely on distributed and collaborative intelligence, ensuring trust and reliability of the underlying Machine Learning (ML) models becomes critical, especially when raw data cannot be centrally aggregated due to privacy or security concerns. Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed data sources while preserving data privacy. However, the inherent opacity of the FL process introduces several challenges; individual data contributions, as well as local node updates, remain inaccessible to centralized oversight, hindering the overall trustworthiness of the global model training process. In addition, data and updates of the local model can be biased, or even maliciously modified, negatively affecting the global model. In the context of DTs, such vulnerabilities can directly compromise decision-making, leading to operational safety risks and affecting the system’s reliability. To this end, we propose TrustFLow, a comprehensive framework that integrates Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs) deployed over a distributed ledger infrastructure to securely trace the provenance of data and evolution of models in FL. TrustFLow tracking capabilities enable linking data to sources, tamper-proof monitoring, and process traceability. In addition, TrustFLow provides crucial functionalities for estimating the influence of individual producers (datasets) on the global model and for revoking both biased data and the global model(s) influenced by them. These features contribute to a broader vision of federated governance, where accountability, transparency, and trust are enforced between all participants. In addition, our framework offers the benefit of seamless integration into any FL-driven DT architecture. To evaluate our proposal, we have conducted an extensive set of experiments that measure the efficiency and effectiveness of the framework under different settings.