Smart Grid Worm Detection Based on Deep Reinforcement Learning

Abstract

Listen

Translate article

Worm detection in smart grids usually depends on monitoring and analyzing the abnormal change of network traffic features such as spectrum features. However, a class of worms with traffic morphing technique is able to effectively change the propagation speed, control the network traffic features and thus evade detection. In this paper, we propose a deep reinforcement learning based worm detection scheme for smart grids which choose the worm detection threshold to improve the detection performance based on the state consisting of the traffic logs, the number of infected SMs and the detection accuracy without knowing the worm propagation model. This scheme designs two neural networks with fully connected layers to reduce computational complexity and avoid the overestimation of the worm threshold selection policy. Simulation results show that our proposed scheme significantly decreases the detection delay, improve the detection accuracy and increases the utility of the control center.