The field of graph neural networks (GNNs) is moving towards improving security and reliability, with a focus on detecting and preventing various types of attacks, such as model extraction attacks, graph poisoning attacks, and adversarial attacks. Researchers are proposing novel frameworks and methods to enhance the robustness of GNNs, including real-time detection of evolving attack patterns, incorporating temporal dynamics into message passing, and using gradient representations and deceptive predictions to protect against data-free model stealing attacks. Additionally, there is a growing emphasis on developing certified defenses against arbitrary perturbations, with deterministic robustness guarantees. Noteworthy papers in this area include ATOM, a real-time MEA detection framework, and TeMP-TraG, a novel graph neural network mechanism that incorporates temporal dynamics into message passing. Model-Guardian, a defense framework against data-free model stealing attacks, is also a significant contribution. Furthermore, the proposal of PGNNCert, the first certified defense of GNNs against poisoning attacks under arbitrary perturbations with deterministic robustness guarantees, marks an important milestone in this field.
Advancements in Graph Neural Network Security and Reliability
Sources
ATOM: A Framework of Detecting Query-Based Model Extraction Attacks for Graph Neural Networks
TeMP-TraG: Edge-based Temporal Message Passing in Transaction Graphs
Robustness of deep learning classification to adversarial input on GPUs: asynchronous parallel accumulation is a source of vulnerability
Jailbreaking the Non-Transferable Barrier via Test-Time Data Disguising
Model-Guardian: Protecting against Data-Free Model Stealing Using Gradient Representations and Deceptive Predictions
Enhance GNNs with Reliable Confidence Estimation via Adversarial Calibration Learning
Deterministic Certification of Graph Neural Networks against Graph Poisoning Attacks with Arbitrary Perturbations
Activation Functions Considered Harmful: Recovering Neural Network Weights through Controlled Channels
Efficient Adversarial Detection Frameworks for Vehicle-to-Microgrid Services in Edge Computing
Are We There Yet? Unraveling the State-of-the-Art Graph Network Intrusion Detection Systems
$\beta$-GNN: A Robust Ensemble Approach Against Graph Structure Perturbation