The Design and Implementation of Secure Distributed Image Classification Model Training System for Heterogenous Edge Computing

Deep learning provides many new and efficient solutions for edge computing. We study training image classification models on edge devices in this paper. Although there have been many researches on deep learning in edge computing. Most of them did not consider the impact of the limited service capabilities of edge devices, the problem of straggler and insecurity of training data on the system. We design a new distributed computing system to train image classification models on edge devices. To be more specific, we vectorize the convolutional neural network (CNN) to transform it to a lot of matrix multiplications. These matrix multiplications can be arbitrarily cut into many smaller matrix multiplications suitable for computing on edge devices. Besides, our system utilizes codes to ensure the stability and security of distributed matrix multiplications on edge devices. In the performance evaluation, we test the performance of matrix multiplications and a CNN model training in our system with uncoded and coded strategies. The evaluation results show that the system with code strategies perform better than with uncoded strategies on the edge devices having the problem of straggler. In summary, we design a secure distributed image classification model training system for heterogenous edge computing.