Trojan Detection using Convolutional Neural Network

Abstract

As machine learning becomes more popular and computing power is increasingly needed, hardware-optimized neural networks and other learning models are increasingly needed. In the course of technology's evolution, machine learning and artificial intelligence will also likely be well trained in the near term. The modern fabulous production hardware business model leads unfortunately to security deficiencies throughout the supply chain and to economics. In this article, these safety problems are emphasized through the introduction of Trojan hardware attacks on neural networks to expand the current neural network security taxonomy. This paper proposes the development of a new framework to insert malicious trojans into a classifier application for the neural network. An algorithm using a convolutional neural network is used to evaluate the ability, if this algorithm adds 0.03 percent trojan, it can effectively classify an input gauge as a cluster in any convolution neural network with seven layers. Finally, this work is about the potential defense against hardware Trojan attacks to protect neural networks.