CFedDT: Cross-Domain Federated Learning in Digital Twins for Metaverse Consumer Electronic Products

With the development of Consumer Electronic Products (CEP) in the Computer, Communication, and Consumer electronics (3C) industry, research over high speed, high precision, and reliability CEP has been an urgent need. Due to the high manufacturing requirements, it has been challenging to product, transmit, assembly, or test the CEP. Metaverse with Virtual Reality (VR)/Augmented Reality (AR) technologies connect reality and virtual, which provides a more convenient manufacturing platform for CEP, thus bringing up the concept of Metaverse Consumer Electronic Products (MCEP). Digital Twins (DT) has effectively provided an emulated software replica to create the Metaverse. DT for MCEP (DTM) facilitates the implementation of more advanced applications, especially Artificial Intelligence (AI) applications. An essential component in DTM AI applications is the labeled dataset, but such carefully labeled data are usually inaccessible since labeling is costly. In addition, the unlabeled data are usually stored locally, referred to as Federated Learning (FL). Previous works adopted Domain Adaptation to construct a pseudo-labeled dataset. However, those approaches have rising potential in accuracy, convergence speed, and communication efficiency. In this paper, we propose cFedDT, a cross-domain FL algorithm, to construct a pseudo-labeled dataset in DTM. cFedDT contains three modules. Ensembled Knowledge for Cross Distillation distills the knowledge from the source models to the target model to improve the target accuracy. Adaptive Batch adjustment for Stable Convergence gradually adjusts the number of batches on each client to speed up the converging process. Magnitude Pruning for Efficient Communication reduces the parameters that require communication to speed up the training. Our experimental results illustrate the priority of cFedDT. Compared with other state-of-the-art algorithms, cFedDT can increase the target accuracy by 9.8% or reduce up to 99.97% of the communication with a target accuracy loss of less than 2.7%.