OblivDGL: Oblivious Deep Graph Learning Made Practical

What it claims

OblivDGL is a secure GNN training and inference framework from Yiming Qin, Xiaoning Liu, Shangqi Lai, Joseph K. Liu and Xingliang Yuan (contact at Monash). The authors argue that prior secure GNN protocols fail on three counts at once:

• They leak some graph information.
• They offer limited support for secure training.
• They scale poorly to large graphs.

OblivDGL's stated privacy guarantee is that it reveals nothing except the total node and edge counts. The authors call this the most stringent guarantee among prior secure GNN systems.

How it works

The abstract gives four design steps:

1. It distills the secure computation patterns of GNNs into one generalized form, modelled on Deep Graph Library (DGL).
2. An oblivious message passing paradigm runs over that form for both training and inference.
3. A new oblivious graph representation implements the paradigm. It is built on function secret sharing (FSS), which the authors pick for its low latency.
4. Tailored parallelization designs let the system process encrypted graphs at scale. The authors name this as a standout feature.

Evidence

The comparison is against "state-of-the-art secure GNN frameworks" on four benchmark graph datasets. The abstract names neither the baselines nor the datasets.

All figures are maxima. Graph sizes, hardware and network settings are not given.

Limits

The page holds only the abstract and metadata. It does not give:

• the threat model, or the number of non-colluding servers that FSS normally assumes;
• the GNN architectures supported;
• accuracy relative to plaintext training;
• any code release.

The paper's status is listed as "Published elsewhere. Major revision. NDSS 2027." That means it is under major revision for NDSS 2027 and has not been accepted, so results may change. The ePrint version was received on 27 October 2026 and approved on 28 October 2026.

The paper is licensed CC BY-NC-ND. This restricts commercial reuse and derivative works of the text.