Secure Multi-Party Computation in Federated Learning for Industrial IoT

Abstract

Industrial Internet of Things (IIoT) is reshaping the production line and infrastructure-grade applications by making their operations smart using distributed sensor networks and embedded computing to realize intelligent automation, real-time analytics, and predictive maintenance. But with machine learning, it is becoming one of the components in such systems, and the security of the sensitive data used in the operation becomes of vital necessity. Federated Learning (FL) has been suggested as an effective method to conduct collaborative training of models on distant IIoT devices devoid of centralization of uncooked information, which thus retains local confidentiality. Nevertheless, this does not eliminate the privacy risks models face, including model inversion and gradient leakage, which are still a threat to most conventional FL systems, and in adversarial contexts. In order to bridge those vulnerabilities, this paper proposes a new privacy-preserving FL design that incorporates the Secure Multi-Party Computation (SMPC) into the model aggregate process. In the suggested framework, many IIoT nodes can jointly compute the encrypted model update using additive secret sharing scheme to achieve the effect that neither the node nor the aggregator can access to the raw update or the proprietary data. This solves them specifically towards low-power, resource-constrained IIoT edge devices and, to guarantee that they can be computed in such a low-resource environment, applies optimization techniques including model quantization and lightweight cryptographic operations. To compare the system, we test it on several industrial datasets, such as Industry-MNIST, UCI Gas Sensor Array, and NASA C-MAPSS and observe the performance by factors such as model accuracy, system latency, communication overhead, and data leakage attack resilience. As results in our experiments demonstrate, our SMPC-enhanced FL system provides competitive accuracy levels with less than 1 percent accuracy loss compared to regular FL whilst offering much better privacy guarantees and preserving the ability to perform inference within real-time. In addition, the framework can easily be scaled to different numbers of IIoT nodes and can tolerate node dropout and malicious bahavior. The study offers a safe, effective, and expandable platform of implementing collaborative AI models in IIoTs, which opens the path toward reliable industrial intelligence without negatively affecting data security and the functioning of the system.