Analysis of Techniques for Mapping Convolutional Neural Networks onto Cloud Edge Architectures Using SplitFed Learning Method

The Convolutional Neural Network is a machine learning algorithm of increasing interest in recent years for its use in computer vision. Today, there are a lot of applications in safe driving, object recognition, person identification and in healthcare. On the other hand, many devices do not have the computational power to support a deep neural network and, moreover, a machine learning algorithm requires a training set of considerable size for optimization, which is continuously updated and common to multiple users. The shared data relating to the images, can generate security problems to the system by falling within the field of data privacy. Local regulations, such as the GDPR in Europe, provide for high levels of security, in particular data defined “sensitive”, such as biometrics and health data. Using this data in a shared environment can lead to a data breach, not sharing it degrades CNN’s performance. In this article we will illustrate the mechanisms of subdivision of a convolutional neural network between edge devices, with limited computational power and a public cloud platform. The distribution of the computation aims at convolution of the neural network and at preserving system security. In particular, an example of distribution will be illustrated using the tree-computation pattern on SplitFed Learning architecture.