Collaborative Edge-Cloud AI for IoT Driven Secure Healthcare System

In healthcare applications like monitoring patients in ICUs and performing precision robotic surgeries, IoT and sensor networks have become indispensable. These sensors generate a large amount of data that, when processed and visually presented to a medical professional, assists in the more accurate diagnosis and treatment of ailments. For some time now, hospital administrations have been taking advantage of public cloud(referred to as main clouds in this paper) resources to store and process patient data using the advanced AI analytical tools that these clouds provide. However, taking all the medical sensor data to the main cloud encounters network congestion and latencies that may negatively impact the outcomes. In this situation the power of edge-AI may appear appealing, but the state-of-the-art does not allow all the tasks of training complex AI models and drawing inference from them to take place at the edge. Techniques of complexity reduction like pruning and quantization have been applied to reduce storage and processing burden, but they compromise accuracy of the models. Researchers now agree on the necessity of collaborative edge-main cloud AI for demanding workloads.It is, however, necessary to realize that the multi-layer IoT-Edge-Main Cloud arrangement has an expanded attack surface. Any malicious attack on the dataflows among various layers may threaten patients’ quality of life or even their lives. Although AI can be used to secure these dataflows, using large neural network models centrally on the main cloud results in long training and inference dispersion times. We propose a collaborative, hierarchically merged technique to help train large neural network models in real-time. This is achieved by synthesizing the main cloud model using the trained layers of the edge models, resulting in a dramatic reduction in the training times of the model in the main cloud while achieving high detection accuracy. As we shall see in the description, this method removes some of the problems faced with other collaborative methods, like federated learning, which works by disaggregating models for sharing training load.