Towards Transmission-Friendly and Robust CNN Models Over Cloud and Device

Deploying deep convolutional neural network (CNN) models on ubiquitous Internet of Things (IoT) devices has attracted much attention from industry and academia since it greatly facilitates our lives by providing various rapid-response services. Due to the limited resources IoT devices, cloud-assisted training CNN become mainstream. However, most existing related works suffer a large amount model parameter transmission weak robustness. To this end, paper proposes a framework with low strong robustness. In proposed framework, we first introduce MonoCNN, which contains only few learnable filters, other filters are nonlearnable. These nonlearnable filter parameters generated according certain rules, i.e., generation function (FGF), can be saved reproduced random seeds. Thus, cloud server needs send these seeds device. Compared transmitting all parameters, sending several significantly reduce transmission. Then, investigate multiple FGFs enable device use FGF generate combine them into MonoCNN. MonoCNN is affected not data but also FGF. The rules play role in regularizing thereby improving its Experimental results show that compared state-of-the-art methods, transfer between while performance approximately 2.2% when dealing corrupted data.