Federated Learning Approaches for Privacy Preserving AI Applications

Abstract

Federated Learning (FL) enables collaborative model training without centralizing data, which is ideal for privacy-preserving AI applications. Concurrently, Digital Twins (DTs)—realistic virtual replicas of physical manufacturing systems—are central to intelligent manufacturing. Integrating FL with DTs offers a compelling framework for preserving data privacy while maintaining high-fidelity twin behavior and insight.

In this work, we propose a unified Federated Digital Twin (Fed-DT) architecture for intelligent manufacturing systems, enabling decentralized model training across factory sites while ensuring privacy, robustness, and adaptability. At the core is a DT-assisted knowledge distillation framework: a DT located in the server trains a resource-rich “teacher” model; individual edge clients (factories) train lightweight “student” models locally, guided via knowledge distillation and resource allocation optimized through reinforcement learning. This approach supports heterogeneous model architectures tailored to device capability and ensures scalable, private learning.

We further explore privacy and robustness by integrating secure client clustering and offloading strategies—where high-resource clients assist weaker ones—within a clustered FL (CISCO-FL) framework. The hybrid model addresses challenges of model heterogeneity, client selection, limited IoT resources, and communication overhead.

Simulations and case studies in additive manufacturing scenarios (e.g., 3D printing) show improvements in model accuracy (average gain ~5 pp), reduced delay, and increased convergence rates under non-IID data distributions. We observe stability across heterogeneous DT deployments, demonstrating Fed-DT’s feasibility in real-world industrial settings.