Reinforcement learning based a non-zero-sum game for secure transmission against smart jamming

Abstract Smart jammer and smart anti-jammer have always been attacked and defensed in a contradictory way. In fact, there exists fundamental trade-off between all evolved parties. It is well known that only through interactive training with powerful opponents can the strategy optimization ability in actual combat be improved. In the process of electronic countermeasures between the communication system and the jammer, the traditional electronic warfare attacker uses an open-loop jamming method, that is, the opponent's information cannot be obtained after the jamming, which greatly reduces the combat effect. In order to improve the strategic optimization capabilities of the combatant, cognition and intelligence are introduced into electronic countermeasures. In this paper, we start from modeling a non-zero-sum game, and analyze the Nash equilibrium (NE) of the static secure game and the conditions for its existence. Then, we design a multi-agent reinforcement learning framework with a optimal power control strategy in the dynamic game between the smart jammer and the trained base station (BS). Finally, due to the non-cooperative hostile relationship between the two sides in the actual combat, we add the eavesdropping function and the jamming effect evaluation modular to build a cognitive closed-loop. The experiment shows that the intelligent jammer with the eavesdropping function can seriously reduce the performance of the interfered communication party. However, the intelligent BS after training can effectively combat smart jamming. It can be demonstrated that confrontation training can improve the intelligence level of agents.