Personalized Federated Learning for Generative AI Empowered Digital Twin Networks

Generative artificial intelligence (GenAI) has emerged as a fundamental building block for Digital Twin Networks (DTNs) due to its advantage in synthesizing realistic data and modeling complex system behaviors in real time. However, training high-fidelity GenAI models for DTN while fulfilling latency constraints is challenging, and the presence of data heterogeneity further complicates this task. To address this challenge, we propose a novel framework, termed Personalized Federated Generative Adversarial Network (PFGAN), which trains a high-fidelity model tailored to each client rather than relying on a shared global model. To improve model fidelity under data heterogeneity, we formulate an optimization problem that minimizes both training and generalization errors while satisfying the latency constraint. To make the problem tractable, we theoretically derive the upper bounds of both the training error and generalization error of PFGAN to approximate the objective function and decompose it into two subproblems: optimal weight estimation and aggregation frequency optimization. Subsequently, a closed-form solution of the optimal weight estimation problem is derived to compute aggregation weights that yield personalized models, and an efficient aggregation frequency optimization method is proposed to minimize the training error under the latency constraint. Experimental results on three datasets show that PFGAN generates more accurate data and achieves higher training efficiency than the baselines.