Leveraging Graph Clustering for Differentially Private Graph Neural Networks
Bibliographic record
Abstract
Differential Privacy is a standard approach for ensuring privacy in deep learning models, but its effectiveness is less certain when applied to message-passing Graph Neural Networks (GNNs). GNNs, which process graph-structured data, generate node representations by aggregating information from neighboring nodes. At the k-th layer, GNNs propagate information across a node’s k-hop neighborhood, causing interconnected nodes to influence each other’s representations. Consequently, protecting the privacy of a single node, edge, or feature often requires safeguarding information about related graph elements as well. Previous methods have addressed this issue by injecting noise into the data, though finding the right balance is challenging; while more noise increases privacy, excessive noise degrades model output. Other approaches use custom architectures to decouple neighborhood aggregation from node representation learning, but these solutions often struggle to scale to large graphs. We propose a strategy for training subgraph-level DP-GNNs by extracting disjoint subgraphs from the training dataset and applying the DP-SGD algorithm, treating each subgraph as an independent sample. This method protects the privacy of target nodes and their neighbors. Our graph partitioning approach is inspired by community detection techniques, which help preserve relevant connections within partitions. By restructuring the training data, our solution enhances privacy protection while maintaining model utility, ultimately outperforming existing techniques.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.014 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.005 |
| Open science | 0.003 | 0.003 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".