DESIGN: Encrypted GNN Inference via Server-Side Input Graph Pruning

ICLR 2026 Conference Submission15372 Authors

Keywords: GNN, FHE, Pruning, Encrypted inference

Abstract: Graph Neural Networks (GNNs) enable powerful graph learning, yet Fully Homomorphic Encryption (FHE) makes inference prohibitively expensive. We present DESIGN (EncrypteD GNN Inference via sErver-Side Input Graph pruNing), a server-side framework that reduces FHE cost without client changes. DESIGN computes encrypted degree–based importance scores and uses homomorphic comparisons to produce multi-level masks, which drive two optimizations: logical pruning of low-importance nodes and edges, and importance-aware assignment of low-degree polynomial activations to most nodes while reserving higher degrees for critical ones. Across standard benchmarks, DESIGN substantially accelerates encrypted GNN inference while maintaining competitive accuracy. Code is available at \href{https://anonymous.4open.science/r/DESIGN-7F93}{\url{https://anonymous.4open.science/r/DESIGN-7F93}}.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 15372