Federated Learning for Microwave Filter Behavior Prediction

Deep learning (DL) technologies have been widely investigated to improve the performance of microwave device behavior prediction. Advanced microwave-related DL technologies utilize independent computers to collect data from the electronic design automation (EDA) software. However, it is essential to note that DL requires a vast amount of high-quality training data. Collecting these data from exact simulation meticulous optimization in EDA is exceptionally time-consuming and computationally intensive. A straightforward way to speed up the process is by collecting quality data from distributed radio frequency (RF) designers. However, this approach may not always be feasible due to the need to maintain the confidentiality of sensitive microwave design information. In this letter, we proposed a federated learning (FL) framework for corporately training DL models for microwave filter behavior prediction. The FL framework aggregates knowledge from various designers without sharing their raw data. The primary experimental results demonstrate the feasibility of the proposed encrypted FL framework for microwave filter applications with superior accuracy and speed.