Study On Security Techniques In Water Communication Networks Based On Machine Learning

Water communication networks include the underwater sensor networks and wireless networks involved with security-sensitive information,such as the location of sensor nodes and ships,and are vulnerable to spoofing and jamming attacks,which can tamper the network data and reduce the lifetime of sensors and further result in the communication interruption and denial of service attack.Underwater sensors with constrained computational resource and battery life cannot support security techniques with high algorithm and communication complexity.Therefore,we investigate the machine learning based security techniques with reduced overhead for water communication networks.In this thesis,a reinforcement learning(RL)based physical(PHY)-layer authentication is investigated for underwater sensor networks without knowledge of underwater channel model and spoofing attack model,in which the sink node selects the optimal authentication parameter with RL technique to improve the authentication accuracy.Deep Q-networks(DQN)-based authentication for sinks with sufficient computing resource further accelerates the learning speed and reduces the authentication error rate with compressed state space.Experiment results validate the DQN-based authentication reduces the detection error rate and increases the network utility.For example,the DQN-based scheme respectively decreases the miss detection rate and false alarm rate by 87.1%and 74.5%,and increases the network utility by 88.1%compared with the benchmark scheme based on the fixed strategy.An online machine learning based PHY-layer authentication system is designed for shipborne wireless network with multiple landmarks,which uses landmarks each with multiple antennas to improve the authentication accuracy with the enhanced spatial resolution of surface station.Without knowing the channel model and attack model,we propose the logistic regression based authentication with higher applicability Coefficient estimation in the authentication model is optimized by the incremental aggregated gradient(IAG)algorithm.Experimental and simulation results verify the IAG-based authentication reduces the detection error rate and computation overhead.For example,the miss detection rate and the false alarm rate are lower than 5×10-5 with the computation overhead 91.3%less than the logistic regression based authentication with Frank Wolfe,with 6 landmarks each with 6 antennas.Meanwhile,the detection error rate of the logistic regression based authentication with IAG is reduced more than 99.0%compared with the hypothesis test based authentication.An anti-jamming transmission framework for underwater sensor networks is proposed to resist j amming,in which reinforcement learning is applied to determine the transmitter's power and receiver's mobility.Without knowing the underwater channel model and jamming model,the DQN-based anti-jamming strategy determines the transmitter's power and receiver's mobility.Experiment results verify that the proposed scheme can improve the underwater transmission performance.For instance,in the DQN-based power control,the SINR is 43.3%higher and the BER is 76.6%lower than the Q-learning strategy.