Particle Swarm Optimized Federated Learning For Securing IoT Devices

Abstract

Federated learning (FL) focuses on interpreting optimization, privacy, and communication but pays little consideration to enhance training and results on the edge devices. The major challenge on these Internet of Things (IoT) devices is efficient training and inference. Another considerable challenge is securing IoT devices for a long time. This paper resolves it by selecting appropriate parameters for building a local machine learning or deep learning (ML/DL) model. Appropriate parameters will make the model's training less computationally expensive and secure the edge or IoT device. So, we propose a particle swarm optimization (PSO) method to optimize the hyper-parameter environments for the bounded DL model in an FL environment. First, we select the 2-gram represented Application Programming Interface (API) calls of the malicious and benign instances for the dataset's feature. Then, API calls of the sample are represented using 2-gram, and their frequency fills the dataset's rows. Later, we represent the sample's feature in a grayscale image and apply the LeNet-5 model. Our experiment indicates that PSO efficiently tunes the hyperparameters of LeNet-5 compared to the grid search method. The near-optimal parameters for FL do not affect the model's accuracy.