Adaptive Noise Trap of Layered Differential Privacy Against Privacy Leakage for Large-Scale Industrial Federated Learning

Benefited from privacy improvement during cooperative training, federated Learning (FL) has been popularly applied for intelligent manufacturing within industrial internet of things (IIoT). However, recent studies have shown the vulnerability of existing FL, since the privacy provided by FL might be inferenced on the wireless training stage by interactive data. More seriously, for industrial applications, traditional FL methods of protecting privacy might lead to a drastic decrease in accuracy, caused by the increasing difficulty and costs of managing massive clients. Thus, this paper proposes an Adaptive Noise Trap (ANT). As a protection strategy for privacy applied in FL, our additive noise defense strategy is able to address the challenges posed by massive privacy in the data level and reduce its defense cost by adding noise defense on the server-side global model in FL. Meanwhile, ANT adjusts the noise size according to the privacy carried in the model, which ensures the privacy defense performance and model accuracy in large-scale IIoT. Especially, network traffic classification is chosen as a typical application scenario for the methodology verifications based on simulations. Through experimental comparisons, ANT achieves a lightweight defense, which uses less noise to achieve the desired defense effect. This lightweight defense can play an unimaginable role in the massive scale of IIoT. Moreover, the simulation results also show that ANT guarantees convergence, accuracy and reliability of federated learning.