Secure Energy-Efficient Implementation of CNN on FPGAs for Accuracy Dependent Real Time Task Processing

A key objective of the fourth industrial revolution or Industry 4.0 is to use processing resources that are reconfigurable and flexible, having additional capability to run smart real time intelligent applications in short time. For this, designers deploy reconfigurable hardware or field programmable gate arrays (FPGAs) in several critical infrastructures, along with cloud and edge platforms. Among the commonly used artificial neural networks, convolutional neural networks (CNNs) is one of the most widely used for various smart applications. Existing strategies of CNN implementation on FPGAs incur significant resources and power, and hence, are not energy efficient. These are even prone to side channel attacks. Moreover, they are associated with prunning and hence, do not generate accurate results, which are important for some real time applications. In our proposed methodology, we pre-compute various operations, mainly operations associated with secret information, which in the present case are weights and bias of the CNN for a particular application and store in available embedded memory blocks (EMBs) of an FPGA. On demand, these pre-computed results are accessed for real time operations. Via this methodology, we eradicate the side channel attacks that steals the weights and biases, obtain low resource utilization, low latency, low power and better energy efficiency, without any loss of accuracy. Such a mechanism is particularly suitable for high accuracy real time smart intelligent applications.