Distributing DNN Training over IoT Edge Devices based on Transfer Learning

Abstract In this paper, an approach for distributing the deep neural network (DNN) training onto IoT edge devices is proposed. The approach results in protecting data privacy on the edge devices and decreasing the load on cloud servers. In addition, the technique may reduce the communication traffic between the cloud and the edge devices. Since the available resources in the edge devices are limited, in the proposed approach, we suggest a heuristic technique for generating a smaller network based on the main network in cloud. Next, by exploiting the knowledge distillation method, the knowledge of the main network is transferred to the generated small network. In this approach, small networks on the edge devices under different datasets are trained where some of their parameters are aggregated for updating the main network parameters on the cloud. The effectiveness of this approach is assessed with some state-of-the-art neural networks. Results show that the approach, the price of preserving the data privacy, is, on average, about 3.5% accuracy loss compared to the case when the network is trained on the cloud and all the datasets of the edge devices are available for training.