| Graph Neural Networks(GNNs)aim to extend the convolutional networks from two-dimensional computer vision to irregular data structures.Recent research has focused on utilizing large-scale datasets,more expressive architectures,and designing superior models through self-supervised learning.To achieve better fitting capabilities,these GNNs have a large number of parameters,which results in high computational costs.Recent work has also emerged in the field of lightweight and resource-efficient GNNs,enabling high performance under computational and memory constraints.Existing methods primarily employ knowledge distillation(KD)to transfer knowledge from high-performance but resource-intensive teacher models to resource-efficient students.Model extraction methods are built upon the foundation of knowledge distillation techniques.However,while knowledge distillation typically retains full knowledge of the target models training data and weights,model extraction does not have access to the target models weights and training data.This study aims to address scenarios where training data is unavailable due to privacy,legality,and security concerns.It utilizes knowledge distillation techniques to extract the knowledge from pre-trained models in both white-box and black-box scenarios,generating high-quality and diverse data to improve the generalization ability of student models.The specific research content and achievements are outlined as follows:(1)Data-Free Adversarial Knowledge Distillation based on Graph Neural NetworksTraditional knowledge distillation relies on the transfer of knowledge between models using the original training data.However,in many cases,the original data may be inaccessible due to copyright or confidentiality issues.For example,in social networks,a users friend list,personal profile information,likes,and comments may be private.To address this issue,this study proposes a Data-Free Adversarial Knowledge Distillation framework for GNNs(DFAD-GNN)based on graph-structured data.Specifically,DFAD-GNN employs a generative adversarial network,consisting of two discriminators representing the pre-trained teacher model and the student model,and a generator that generates synthetic graphs to transfer knowledge from the teacher model to the student model.Extensive experiments on various benchmark models and representative datasets demonstrate that DFAD-GNN significantly outperforms state-of-the-art data-free GNN distillation methods.(2)Data-Free Model Extraction Attacks on Graph Neural Networks Most model extraction attacks require substitute data samples from the real world as queries to steal the target model,which is impractical in practical applications.To bridge this gap,this study proposes a data-free model extraction attack framework called StealGNN,specifically designed for GNNs.StealGNN improves existing GNN extraction attacks in three major aspects:1)it is entirely data-free,as it does not require any real node features or graph structures to extract the GNN model;2)it can be instantiated for both node classification and link prediction tasks,posing a significant intellectual property threat to various graph tasks;3)it can handle the most challenging hard-label attack scenario,where the attacker has no knowledge of the target GNN model and can only obtain prediction labels by querying the victim model.Experimental results on node classification benchmark datasets demonstrate that StealGNN can effectively extract knowledge from GNNs. |
