TY - JOUR
T1 - DIFFERENTIAL PRIVACY GUARANTEES FOR ANALYTICS AND MACHINE LEARNING ON GRAPHS
T2 - A SURVEY OF RESULTS
AU - Mueller, Tamara T.
AU - Usynin, Dmitrii
AU - Paetzold, Johannes C.
AU - Braren, Rickmer
AU - Rueckert, Daniel
AU - Kaissis, Georgios
N1 - Publisher Copyright:
© T.T. Mueller, D. Usynin, J.C. Paetzold, R. Braren, D. Rueckert, and G. Kaissis Creative Commons (CC BY-NC-ND 4.0).
PY - 2024
Y1 - 2024
N2 - We study differential privacy (DP) in the context of graph-structured data and discuss its formulations and applications to the publication of graphs and their associated statistics, graph generation methods, and machine learning on graph-based data, including graph neural networks (GNNs). Interpreting DP guarantees in the context of graphstructured data can be challenging, as individual data points are interconnected (often non-linearly or sparsely). This differentiates graph databases from tabular databases, which are usually used in DP, and complicates related concepts like the derivation of per-sample gradients in GNNs. The problem is exacerbated by an absence of a single, well-established formulation of DP in graph settings. A lack of prior systematisation work motivated us to study graph-based learning from a privacy perspective. In this work, we systematise different formulations of DP on graphs, and discuss challenges and promising applications, including the GNN domain. We compare and separate works into methods that privately estimate graph data (either by statistical analysis or using GNNs), and methods that aim at generating new graph data. We conclude our work with a discussion of open questions and potential directions for further research in this area.
AB - We study differential privacy (DP) in the context of graph-structured data and discuss its formulations and applications to the publication of graphs and their associated statistics, graph generation methods, and machine learning on graph-based data, including graph neural networks (GNNs). Interpreting DP guarantees in the context of graphstructured data can be challenging, as individual data points are interconnected (often non-linearly or sparsely). This differentiates graph databases from tabular databases, which are usually used in DP, and complicates related concepts like the derivation of per-sample gradients in GNNs. The problem is exacerbated by an absence of a single, well-established formulation of DP in graph settings. A lack of prior systematisation work motivated us to study graph-based learning from a privacy perspective. In this work, we systematise different formulations of DP on graphs, and discuss challenges and promising applications, including the GNN domain. We compare and separate works into methods that privately estimate graph data (either by statistical analysis or using GNNs), and methods that aim at generating new graph data. We conclude our work with a discussion of open questions and potential directions for further research in this area.
KW - differential privacy
KW - graph analytics
KW - graph neural networks
KW - graph-structured data
UR - http://www.scopus.com/inward/record.url?scp=85185316952&partnerID=8YFLogxK
U2 - 10.29012/jpc.820
DO - 10.29012/jpc.820
M3 - Article
AN - SCOPUS:85185316952
SN - 2575-8527
VL - 14
JO - Journal of Privacy and Confidentiality
JF - Journal of Privacy and Confidentiality
IS - 1
ER -