OblivDGL: Oblivious Deep Graph Learning Made Practical

What it is

OblivDGL is a framework for secure GNN training and inference, by Yiming Qin, Xiaoning Liu, Shangqi Lai, Joseph K. Liu and Xingliang Yuan. The ePrint record (2026/2655, received 27 October 2026, approved 28 October 2026) lists it as a major revision published at NDSS 2027. The target setting is Machine Learning as a Service, where both graph data and proprietary GNN models go to a third-party platform.

How it works

The design follows Deep Graph Library. The authors reduce the secure computation patterns of GNNs to one generalized form and build an oblivious message passing paradigm on it. That paradigm rests on a new oblivious graph representation, realized with function secret sharing (FSS), which the abstract describes as lightweight and low latency. Parallelization designs tailored to encrypted graphs provide the scalability. The abstract gives no protocol detail beyond this.

Results the abstract reports

• Training: up to 1024x lower communication and up to 14.3x faster runtime than state-of-the-art secure GNN frameworks.
• Inference: up to 754x lower communication and a 4.19x speedup.
• Evaluation covers four benchmark graph datasets. The abstract does not name them or the baseline systems.

Caveats

The authors say prior secure GNN designs leak some graph information, support secure training only in a limited way, and scale poorly. OblivDGL reveals only the total number of nodes and edges. The abstract does not state the threat model, the number of parties, or the absolute graph sizes tested. The figures are best-case ("up to") numbers, so check the full paper for per-dataset results before relying on them.