Ameliorating Federated Learning using Dynamic Inertia Weight-Based Advanced Particle Swarm Optimization for Consumer Electronic Devices

The extensive utilization of consumer electronic devices such as smartphones, smart wearables, and smart home technology has resulted in significant surge in data production. Due to storage and data transfer limitations, traditional machine learning techniques are sometimes impracticable for these distributed systems and can cause serious privacy problems. Federated Learning mitigates these issues by maintaining data on local devices. It updates models by consolidating locally trained outcomes on central server, which can be crucial in case of consumer electronic devices. Thus, to tackle these issues, this article presents new method named Dynamic inertia weight-based federated advanced particle swarm optimization (DIW-FedAPSO). It uses dynamic inertia weight strategy in advanced particle swarm optimization to select inertia weight dynamically for providing optimal velocity to consumer electronic devices and transmitting obtained optimal score after performing the local training instead of sending and averaging weights of devices as traditional federated learning method does. The experimental evaluations on different datasets (CelebA, FFHQ) under different non-iid data heterogeneity settings shows that proposed method attains better accuracy, while maintaining data privacy and enhances communication efficiency while minimizing number of communication rounds required to attain targeted accuracy over all datasets than other currently existing methods.