Privacy-Focused Edge AI-on-IoT with Federated Learning

Artificial Intelligence (AI) running locally on the Internet-of-Things (IoT) Edge devices is referred to as Edge AI. Edge AI can be used to augment IoT applications prioritizing data privacy, low latency, and high availability. Federated Learning (FL) is a Machine Learning technique based on the concept of collaborative model training. With FL, a privacy-focused, flexible Edge AI framework can be created for IoT applications. This paper proposes a FL-based Edge AI, using a Convolutional Neural Network (CNN) for image classification. The primary research objective is to create a flexible, privacy-focused Edge AI framework capable of catering to a variety of IoT Edge devices. To improve accessibility of the Edge AI, optimization has been attempted in terms of finding optimal training parameters for FL and implementing model optimization. Our proposed Edge AI framework is able to accommodate several types of IoT Edge devices, and is capable of being trained via FL. Experimental results showed an accuracy improvement of 19.59% from modifying training parameters, although a balance was needed to prevent overfitting. Implementing model optimization reduced computing resource usage during inference. A CPU time reduction of 59.93% was observed, with only a 3.60% corresponding drop in model accuracy. Furthermore, model optimization also enabled a low-power IoT Edge device to run a complex CNN model, a task it was previously incapable of doing.