부산대학교 도서관

Ensuring robustness against adversarial attacks is imperative for Machine Learning (ML) systems within the critical infrastructures of the Industrial Internet of Things (IIoT). This paper addresses vulnerabilities in IIoT systems, particularly in distributed environments like Federated Learning (FL) by presenting a resilient framework - Secure Federated Learning (SFL) specifically designed to mitigate data and model poisoning, as well as Sybil attacks within these networks. Sybil attacks, involving the creation of multiple fake identities, and poisoning attacks significantly compromise the integrity and reliability of ML models in FL environments. Our SFL framework leverages a Digital Twin (DT) as a critical aggregation checkpoint to counteract data and model poisoning attacks in IIoT’s distributed settings. The DT serves as a protective mechanism during the model update aggregation phase, substantially enhancing the system’s resilience. To further secure IIoT infrastructures, SFL employs blockchain-based Non-Fungible Tokens (NFTs) to authenticate participant identities, effectively preventing Sybil attacks by ensuring traceability and accountability among distributed nodes. Experimental evaluation within IIoT scenarios demonstrates that SFL substantially enhances defensive capabilities, maintaining the integrity and robustness of model learning. Comparative results reveal that the SFL framework, when applied to IIoT federated environments, achieves a commendable 97% accuracy, outperforming conventional FL approaches. SFL also demonstrates a remarkable reduction in loss rate, recording just 0.07 compared to the 0.14 loss rate experienced by standard FL systems. These findings highlight the efficiency and applicability of the SFL framework in enhancing data security and traceability within the IIoT ecosystem.