A modified Q-learning algorithm to solve cognitive radio jamming attack

Since the jamming attack is one of the most severe threats in cognitive radio networks, we study how Q-learning can be used to pro-actively avoid jammed channels. However, Q-learning needs a long training period to learn the behaviour of the jammer. We take advantage of wideband spectrum sensing to speed up the learning process and we take advantage of the already learned information to minimise the number of collisions with the jammer. The learned anti-jamming strategy depends on the elected reward strategy which reflects the preferences of the cognitive radio. We start with a reward strategy based on the avoidance of the jammed channels, then we propose an amelioration to minimise the number of frequency switches The effectiveness of our proposal is evaluated in the presence of different jamming strategies and compared to the original Q-learning algorithm. We compare also the anti-jamming strategies related to the two proposed reward strategies.