The Research On Attack And Defense Based On Network Representation Learning

Real-world complex systems can be represented and analyzed as networks.During the past few decades,network science has emerged as an essential interdisciplinary field aiming at using network as a tool to characterize the structure and dynamics of complex systems,including social networks,citation networks,protein networks and transport networks.Quite recently,numerous network embedding have been proposed to map the nodes of a network into vectors in Euclidean space,which largely facilitates the application of machine learning methods in graph data mining.Network embedding solves the problem of high dimensions and sparseness of the original network data.It builds a bridge between machine learning and network science,enabling many machine learning algorithms to be applied in network analysis.Network algorithm security and its related issues are studied in depth and adversarial attack and defense strategies based on network embedding have been proposed.(1)Euclidean distance-based adversarial attack strategy: A new Euclidean distance-based adversarial attack strategy has been proposed.Unlike traditional white-box attack algorithms,this strategy does not target at a specific application or a specific algorithm,but it has attack effect on most algorithms and applications based on random walk.This strategy using the information represented by the network as a feature to generate small perturbations,and it can greatly change the sequence of random walk,so as to change the node representation vector of the network.Then many algorithms that use node representation vectors failed.(2)Adversarial attack algorithm based on coding evolution: A new coding scheme for genetic algorithms has been proposed.Coding pre-trains the network data before generating adversarial sample.The coding strategy adds information such as structural characteristics and node homogeneity.It uses discrete binary coding and a series of adaptation processing.Aiming at the problem of excessive loss of the binary mapping,we first proposed and deduced the matrix factorization by adding the sigmoid function,and adopted the negative sampling method to accelerate the original model.(3)SVD-based graph adversarial defense strategy: Inspired by the adversarial attack and defense in the area of computer vision.A graph adversarial defense strategy based on SVD decomposition has been proposed,which can successfully defense poisoning attack on graph representation.This model reconstructs the feature matrix by retaining critical eigenvalues and their components while filtering the perturbations.Experiments show that with the appropriate selection of the singular value threshold,our proposed SVD decomposition-based graph adversarial defense strategy can significantly improve the prediction accuracy of the graph embedding algorithm under adversarial samples.