Dyn-Backdoor: Backdoor Attack on Dynamic Link Prediction

Dynamic link prediction (DLP) makes graph prediction based on historical information. Since most DLP methods are highly dependent on the training data to achieve satisfying prediction performance, the quality of the training data is crucial. Backdoor attacks induce the DLP methods to make wrong prediction by the malicious training data, i.e., generating a subgraph sequence as the trigger and embedding it to the training data. However, the vulnerability of DLP toward backdoor attacks has not been studied yet. To address the issue, we propose a novel backdoor attack framework on DLP, denoted as Dyn-Backdoor. Specifically, Dyn-Backdoor generates diverse initial-triggers by a generative adversarial network (GAN). Then partial links of the initial-triggers are selected to form a trigger set, according to the gradient information of the attack discriminator in the GAN, so as to reduce the size of triggers and improve the concealment of the attack. Experimental results show that Dyn-Backdoor launches effective backdoor attacks on several state-of-the-art DLP models with a success rate more than 90%. Additionally, we conduct a possible defense against Dyn-Backdoor to testify its resistance in defensive settings, highlighting the needs of defenses for backdoor attacks on DLP.

I. Introduction

Our lives are surrounded by various graphs, which are used to describe complex systems [1], [2], [3], such as social networks [4], [5], biological networks [6], electric systems [7], economics [8], and neural networks [9]. Dynamic link prediction (DLP), inferring the topology of graph in the future based on the historical information of the network, has a wide range of applications, e.g., Taobao, Amazon. Numerous DLP methods have been proposed. For example, similarity based methods [10], [11] redefine common neighbor or resource allocation based on different timestamps obtaining the similarity between nodes on DLP. For random walk based approaches [12], [13], [14], [15], they can reduce the complexity of the model, since they usually take a walk of the local structure. With the success of deep learning, DLP methods based on deep learning [16], [17], [18], [19], [20], [21] are thoroughly studied. They mainly use the nonlinear and hierarchical nature of neural networks to capture the evolving pattern of dynamic networks. In particular, the experiments in [19] and [22] demonstrate that DLP methods based on deep learning generally outperform the non-deep ones.