Reinforcement Learning Based Wireless Network Secure Against Smart Attacks

With the development of smart programming devices,network attackers have the ability of environment awareness and can choose their attack type and mode according to the sensed environment.These attackers are referred to as smart attackers due to their powerful adaptivity and selectivity.Currently,most works on network security assume that the attacker can launch a certain type of attacks,and study the defense mechanisms against this type of attacks.However,little research has been done in smart attackers which have uncertain type of attacks.Compared with the attackers with single type of attacks,smart attackers can launch multiple types of attacks with greater freedom and more pertinence,and thus threaten the network security.In this thesis,the smart attacks in mobile offloading systems are investigated,and a secure mobile offloading scheme is designed.We formulate a mobile offloading game against smart attacks,and Nash and Stackelberg equilibria of the mobile offloading game are provided according to the sequence of actions to reveal the influence of the channel condition on attack type,defense mode and offloading rate.A Q-learning based mobile offloading strategy is proposed,and the simulation results show that the proposed offloading strategy improves the utility of the mobile device and decreases the attack rate effectively.Compared with the random offloading strategy,the average spoofing rate and the average jamming rate of the proposed strategy decrease by 47%and 8%,respectively,if the channel gain is 1.Mobile users have greater freedom with the development of smart devices.For example,users can control the actions of the unmanned aerial vehicle(UAV).Human decisions are usually subjective and may deviate from the result of the expected utility theory.We investigate the smart attacks in UAV systems from a user-centric view,and study the UAV communication strategy against smart attacks.We apply the prospect theory to formulate a subjective game between the UAV and the smart attacker,and reveal the effects of the subjectivity of the UAV operator and the smart attacker on the security of UAV systems.The Deep Q-Network(DQN)based UAV power allocation strategy is proposed.Simulation results show that the DQN-based power allocation scheme improves the secrecy capacity and signal to interference plus noise ratio of the UAV system,and the safe rate of the UAV system with the DQN-based scheme is 7%and 11%higher than that of the Q-learning and the WoLF-PHC based schemes.