Research On Selective Forwarding Attack Detection Scheme Based On Deep Learning

A Wireless Sensor Network(WSN)is a network containing a large number of sensor nodes,routed in a self-organizing,multi-hop manner to transmit information through wireless network.Sensor nodes are responsible for sensing physical quantities and collecting information about the area in which they are located,and then passing the data to sink node(SN)for processing.SN is responsible for sending the data to the end user and has sufficient energy and capacity to handle complex calculations.Because wireless sensor networks use broadcast communication and often deploy nodes in open or unattended areas,they are vulnerable to outside attacks.Selective forwarding attacks,a type of attack,are located at the network layer,which compromises the network with the help of legitimate node identities.Because it causes node packet loss with uncertainty in number and time,it is not easily detected and has become a hot and difficult area of research for attack detection in wireless sensor networks.In thesis,a deep learning-based detection scheme is proposed to deal with selective forwarding attacks,and the main contents are summarized as follows.Firstly,a DBN-based clustered WSN selective forwarding attack detection scheme is proposed.Because selective forwarding attacks are mainly manifested when nodes forward packets,the true identity of a node can be determined by detecting its packet forwarding behavior.The forwarding rate of normal nodes fluctuates slightly under the influence of environment and communication quality,but remains at a high level.On the other hand,malicious nodes will exhibit drastically varying forwarding rates due to their irregular packet loss behavior.To improve the ability to discriminate malicious nodes,thesis employs a deep belief network for detection.Deep Belief Network(DBN)is a probabilistic model that is able to grasp the deeper features of the input data and map them to a lower-order representation.Therefore,with the help of DBN it is possible to find the variation pattern of node forwarding rate and thus distinguish the identity of nodes.The scheme first clusters the nodes to get a dataset suitable for training.Then,the network model is trained and the predicted value of the forwarding rate is obtained.Based on the analysis of the prediction error,malicious nodes are identified.The experimental results show that the scheme can show good detection in the clustered network,and the miss detection rate can reach 0,but the detection period required varies in different malicious proportions.Second,an event-driven WSN selective forwarding attack detection scheme based on DBN is proposed.The scheme is tuned to the data transmission period of event-driven networks and takes into account the low local forwarding rate that may be brought about by harsh environments.Experimental results show that in an ideal environment all malicious nodes can be detected.In the harsh environment,the detection scheme has a miss detection rate and a false detection rate of less than 7%.Furthermore,the short time required to perform a single detection can reduce the damage of the attack as quickly as possible.