A Conditional Diffusion Graph Privacy Protection Model Based on Task-Aware Attention and Hierarchical Adversarial Constraints
Open Access
- 4 June 2026
- journal article
- research article
- Published by MDPI AG in Mathematics
- Vol. 14 (11) , 1988
Abstract
Graph representation learning models inherently capture structural and attribute correlations, which may lead to significant privacy risks under attribute inference attacks. Existing privacy-preserving methods mainly focus on post hoc suppression of generated node representations, while insufficiently considering how to directly weaken sensitive-attribute cues and their inferable associations during the representation recovery stage. To address this issue, this paper proposes a conditional diffusion-based graph privacy protection model with task-aware attention and hierarchical adversarial constraints, termed Task-Aware Adversarial Conditional Diffusion (TA-ACD). Specifically, a Graph Convolutional Network (GCN) is employed to construct a conditional diffusion denoising network for node representation recovery. Sensitive-attribute condition switching is introduced to alter the recovery path, thereby weakening the inferable association between sensitive attributes and node representations. Meanwhile, task-aware attention is embedded into the denoising process to preserve task-relevant discriminative information. Furthermore, node-level and global statistical adversarial constraints are designed to suppress residual sensitive information from both local and global perspectives. Finally, an alternating training strategy is adopted to achieve collaborative optimization between privacy protection and utility preservation. Experimental results on Yale, Rochester, and the Credit defaulter datasets show that TA-ACD achieves a competitive privacy–utility trade-off compared with representative baselines. On the Yale dataset, TA-ACD reduces sensitive-attribute inference accuracy from 86.3% to 62.1%, corresponding to a decrease of 24.2 percentage points, while maintaining a utility-attribute classification accuracy of 85.6%.Keywords
Funding Information
- Guizhou Provincial Science and Technology Department (QianKeHe Basic MS [2025] 622)
- Guizhou Provincial Science and Technology Department (QianKeHe Basic-[2024] Youth 126)
This publication has 23 references indexed in Scilit:
- Object-centric Video Prediction with Mask-guided Spatiotemporal DiffusionMachine Intelligence Research, 2026
- Privacy-Preserving Graph Machine Learning from Data to Computation: A SurveyACM SIGKDD Explorations Newsletter, 2023
- Model Inversion Attacks Against Graph Neural NetworksIEEE Transactions on Knowledge and Data Engineering, 2022
- Adversarial Privacy-Preserving Graph Embedding Against Inference AttackIEEE Internet of Things Journal, 2020
- A Comprehensive Survey on Graph Neural NetworksIEEE Transactions on Neural Networks and Learning Systems, 2020
- Graph neural networks: A review of methods and applicationsAI Open, 2020
- Relational inductive biases, deep learning, and graph networksarXiv, 2018
- Graph embedding techniques, applications, and performance: A surveyKnowledge-Based Systems, 2018
- A Comprehensive Survey of Graph Embedding: Problems, Techniques, and ApplicationsIEEE Transactions on Knowledge and Data Engineering, 2018
- Variational Graph Auto-EncodersarXiv, 2016