pith. sign in

arxiv: 2606.26664 · v1 · pith:OPAUB7VPnew · submitted 2026-06-25 · 💻 cs.CR · cs.AI

TGHE: Template-based Graph Homomorphic Encryption for Privacy-Preserving GNN Inference in Edge-Cloud Systems

Ngoc Bao Anh Le , Thai T. Vu , John Le , Heath Cooper , Jun Shen This is my paper

Pith reviewed 2026-06-26 04:43 UTC · model grok-4.3

classification 💻 cs.CR cs.AI
keywords homomorphic encryptiongraph neural networksprivacy-preserving inferencetemplate fittingfinancial graphsedge-cloud systemsCKKSSIMD packing
0
0 comments X

The pith

Template packing of convergent local tree shapes into shared ciphertexts enables encrypted GNN inference on million-node graphs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing HE-based GNN systems couple per-query cost to global graph size and are limited to small graphs of around 20k nodes. TGHE adopts an ego-centric approach that observes local computation trees in transaction graphs converging to a small set of shapes. It canonicalizes ego-graphs at the edge and packs identical trees into CKKS ciphertexts for SIMD-parallel inference. Approximate Template Fitting and Topology Collapse ensure full SIMD coverage. On the DGraphFin dataset with 3.7M nodes, the Collapse variant delivers 66.9x speedup over sequential encrypted baselines while keeping AUC loss below 0.002.

Core claim

TGHE resolves the scalability barrier of graph-centric HE-GNN systems by exploiting the template phenomenon in transaction graphs, where local computation trees converge into a small set of structural shapes, allowing canonicalization of ego-graphs and packing of structurally identical trees into shared CKKS ciphertexts for SIMD-parallel encrypted inference, with Approximate Template Fitting and Topology Collapse providing full SIMD coverage.

What carries the argument

The template phenomenon, where local computation trees converge to few shapes, combined with Approximate Template Fitting and Topology Collapse to pack identical trees into shared CKKS ciphertexts for SIMD execution.

Load-bearing premise

Local computation trees in transaction graphs converge into a small set of structural shapes that Approximate Template Fitting and Topology Collapse can exploit for full SIMD coverage without meaningful accuracy degradation.

What would settle it

Run the method on a synthetic transaction graph dataset constructed with high structural diversity in local computation trees and measure whether speedup falls below 10x or AUC loss exceeds 0.01.

Figures

Figures reproduced from arXiv: 2606.26664 by Heath Cooper, John Le, Jun Shen, Ngoc Bao Anh Le, Thai T. Vu.

Figure 1
Figure 1. Figure 1: Edge-cloud architecture for encrypted GNN inference. Edge servers [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: From global graph to canonical computation tree. Left: the node’s [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cumulative node coverage vs. template library size ( [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
read the original abstract

Existing homomorphic encryption (HE)-based GNN systems adopt a graph-centric paradigm that couples per-query cost to global graph size, limiting evaluations to at most ~20k nodes and making them incompatible with dynamic, large-scale financial graphs. We propose TGHE (Template-based Graph Homomorphic Encryption), an ego-centric framework that resolves this by exploiting a template phenomenon: local computation trees in transaction graphs converge into a small set of structural shapes. TGHE canonicalizes ego-graphs at the edge and packs structurally identical trees into shared CKKS ciphertexts for SIMD-parallel encrypted inference, with two long-tail optimizers (Approximate Template Fitting and Topology Collapse) ensuring full SIMD coverage. On DGraphFin (3.7M nodes, 4.3M edges), TGHE-Collapse achieves a 66.9x speedup over the sequential encrypted baseline with less than 0.002 AUC loss.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes TGHE, an ego-centric framework for homomorphic encryption (HE)-based GNN inference that exploits an observed 'template phenomenon' in transaction graphs: local computation trees converge into a small set of structural shapes. It canonicalizes ego-graphs at the edge, packs identical trees into shared CKKS ciphertexts for SIMD-parallel encrypted inference, and applies two optimizers (Approximate Template Fitting and Topology Collapse) for full SIMD coverage. The central empirical claim, stated in the abstract, is that TGHE-Collapse achieves a 66.9x speedup over the sequential encrypted baseline on DGraphFin (3.7M nodes, 4.3M edges) with less than 0.002 AUC loss.

Significance. If the empirical result and underlying template assumption hold under scrutiny, the work would meaningfully extend the practical reach of privacy-preserving GNN inference to large-scale financial graphs, addressing the ~20k-node scalability limit of prior graph-centric HE approaches. The ego-centric design and concrete, falsifiable performance numbers on a public large dataset constitute a clear strength; the approach could enable edge-cloud deployments where per-query cost is decoupled from global graph size.

major comments (2)
  1. [Abstract / Experimental Results] Abstract and Experimental Results: The headline claim of 66.9x speedup with <0.002 AUC loss on DGraphFin is presented without implementation details, error analysis, verification of the template convergence phenomenon, description of the sequential encrypted baseline, or any reproducibility artifacts; this directly undermines assessment of the central performance contribution.
  2. [Method] Method section (description of Approximate Template Fitting and Topology Collapse): The core assumption that local computation trees converge to a small set of shapes enabling full SIMD coverage is stated but lacks formal characterization, bounds on approximation error, or empirical validation of the convergence rate; this assumption is load-bearing for both the speedup and the negligible accuracy loss.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential of the ego-centric approach for scaling privacy-preserving GNN inference. We address each major comment below with concrete plans for revision.

read point-by-point responses

  1. Referee: [Abstract / Experimental Results] Abstract and Experimental Results: The headline claim of 66.9x speedup with <0.002 AUC loss on DGraphFin is presented without implementation details, error analysis, verification of the template convergence phenomenon, description of the sequential encrypted baseline, or any reproducibility artifacts; this directly undermines assessment of the central performance contribution.

    Authors: We agree that the manuscript would benefit from expanded experimental documentation. In revision we will add a dedicated 'Experimental Setup' subsection detailing: CKKS parameters (N=2^14, scale=2^40, etc.), hardware (single NVIDIA A100), the sequential baseline (per-ego-graph encryption with no packing or template sharing), and the exact AUC computation protocol. We will include an error breakdown table attributing the <0.002 AUC loss to fitting tolerance and collapse operations, plus a new figure showing template convergence (unique templates vs. sampled ego-graphs on DGraphFin, saturating at ~87 templates). A reproducibility artifact link (anonymized GitHub with preprocessing scripts and parameter files) will be added. These changes directly support the headline numbers. revision: yes

  2. Referee: [Method] Method section (description of Approximate Template Fitting and Topology Collapse): The core assumption that local computation trees converge to a small set of shapes enabling full SIMD coverage is stated but lacks formal characterization, bounds on approximation error, or empirical validation of the convergence rate; this assumption is load-bearing for both the speedup and the negligible accuracy loss.

    Authors: The template phenomenon is presented as an empirical property of transaction graphs rather than a universal theorem. In revision we will add: (i) a formal definition of a template as an isomorphism class of depth-d rooted trees, (ii) empirical convergence plots and rate statistics on DGraphFin demonstrating saturation, and (iii) an explicit error bound for Approximate Template Fitting expressed in terms of the chosen tolerance parameter (maximum structural edit distance). No closed-form theoretical bound on convergence rate across arbitrary graphs will be claimed, as the phenomenon is distribution-dependent; the negligible accuracy loss remains supported by the end-to-end experiments on the target dataset. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central claim is an empirical performance result (66.9x speedup, <0.002 AUC loss on DGraphFin) measured against an external sequential baseline on a public dataset. The method description relies on an observable structural property of transaction graphs (convergence of local computation trees) that is presented as an input premise for the template-based packing and optimizers, not derived from or equivalent to the measured outcome. No equations, fitted parameters renamed as predictions, or self-citation chains reduce the reported result to a tautology by construction. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that transaction-graph ego-trees converge to few structural templates; no free parameters or invented entities are stated in the abstract.

axioms (1)
  • domain assumption Local computation trees in transaction graphs converge into a small set of structural shapes.
    This is the central 'template phenomenon' that enables packing and SIMD efficiency.

pith-pipeline@v0.9.1-grok · 5699 in / 1159 out tokens · 35791 ms · 2026-06-26T04:43:26.066852+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 1 canonical work pages

  1. [1]

    Graph neural networks,

    G. Corso, H. Stark, S. Jegelka, T. Jaakkola, and R. Barzilay, “Graph neural networks,”Nature Reviews Methods Primers, vol. 4, no. 1, p. 17, 2024

    2024

  2. [2]

    A survey on homomorphic encryption schemes: Theory and implementation,

    A. Acar, H. Aksu, A. S. Uluagac, and M. Conti, “A survey on homomorphic encryption schemes: Theory and implementation,”ACM Computing Surveys, vol. 51, no. 4, pp. 1–35, 2018

    2018

  3. [3]

    CryptoGCN: Fast and scalable homomorphically encrypted graph convolutional network inference,

    R. Ran, N. Xu, W. Wang, G. Quan, J. Yin, and W. Wen, “CryptoGCN: Fast and scalable homomorphically encrypted graph convolutional network inference,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 35, 2022. [Online]. Available: https://openreview.net/forum?id=VeQBBm1MmTZ

    2022

  4. [4]

    LinGCN: Structural linearized graph convolutional network for homomorphically encrypted inference,

    H. Peng, R. Ran, Y . Luo, J. Zhao, S. Huang, K. Thorat, T. Geng, C. Wang, X. Xu, W. Wen, and C. Ding, “LinGCN: Structural linearized graph convolutional network for homomorphically encrypted inference,” inAdvances in Neural Information Processing Systems (NeurIPS),

  5. [5]

    Available: https://openreview.net/forum?id=5loV5tVzsY

    [Online]. Available: https://openreview.net/forum?id=5loV5tVzsY

  6. [6]

    Penguin: Parallel-packed homomorphic encryption for fast graph convolutional network inference,

    R. Ran, N. Xu, T. Liu, W. Wang, G. Quan, and W. Wen, “Penguin: Parallel-packed homomorphic encryption for fast graph convolutional network inference,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/2023/hash/ 3cc685788a311fa35d8d41df93e288ca-Abstract-Conference.html

    2023

  7. [7]

    FicGCN: Unveiling the homomorphic encryption efficiency from irregular graph convolutional networks,

    Z. Kan, H. Han, S. Shi, T. Hua, H. Lu, X. Li, J. Mu, and X. Hu, “FicGCN: Unveiling the homomorphic encryption efficiency from irregular graph convolutional networks,” inProceedings of the 42nd International Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 267. PMLR, 2025, pp. 28 832–28 848. [Online]. Available: ht...

    2025

  8. [8]

    torch geometric.datasets.DGraphFin docu- mentation,

    PyTorch Geometric Team, “torch geometric.datasets.DGraphFin docu- mentation,” 2025, accessed via PyG documentation; includes dataset statistics. [Online]. Available: https://pytorch-geometric.readthedocs.io/ en/2.7.0/generated/torch geometric.datasets.DGraphFin.html

    2025

  9. [9]

    DGraph: A large-scale financial dataset for graph anomaly detection,

    X. Huang, Y . Yang, Y . Wang, C. Wang, Z. Zhang, J. Xu, L. Chen, and M. Vazirgiannis, “DGraph: A large-scale financial dataset for graph anomaly detection,” inAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2022. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/ 2022/hash/8f1918f71972789db39...

    2022

  10. [10]

    Yoyo Tricks with AES,

    J. H. Cheon, A. Kim, M. Kim, and Y . Song, “Homomorphic encryption for arithmetic of approximate numbers,” inAdvances in Cryptology – ASIACRYPT 2017, ser. Lecture Notes in Computer Science, vol. 10624. Springer, 2017, pp. 409–437. [Online]. Available: https://link.springer.com/chapter/10.1007/978-3-319-70694-8 15

    work page doi:10.1007/978-3-319-70694-8 2017

  11. [11]

    Stealing links from graph neural networks,

    X. He, J. Jia, M. Backes, N. Z. Gong, and Y . Zhang, “Stealing links from graph neural networks,” in30th USENIX Security Symposium (USENIX Security 21), 2021, pp. 2669–2686. [Online]. Available: https://www.usenix.org/system/files/sec21summer he.pdf

    2021

  12. [12]

    Safeguarding graph neural networks against topology inference attacks,

    J. Fu, Y . Hong, Z. Chen, and W. H. Wang, “Safeguarding graph neural networks against topology inference attacks,” inProceedings of the 32nd ACM Conference on Computer and Communications Security (ACM CCS), 2025, also available as arXiv:2509.05429. [Online]. Available: https://arxiv.org/abs/2509.05429

    arXiv 2025

  13. [13]

    Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial foren- sics,

    M. Weber, G. Domeniconi, J. Chen, D. K. I. Weidele, C. Bellei, T. Robinson, and C. E. Leiserson, “Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial foren- sics,” inKDD Workshop on Anomaly Detection in Finance, 2019, arXiv:1908.02591

    arXiv 2019

  14. [14]

    AMLSim: A multi-agent simulator of anti-money laundering,

    T. Suzumura and H. Kanezashi, “AMLSim: A multi-agent simulator of anti-money laundering,” https://github.com/IBM/AMLSim, 2021

    2021

  15. [15]

    Realistic synthetic financial transactions for anti-money laundering models,

    E. Altman, J. Blanu ˇsa, L. von Niederh ¨ausern, B. Egressy, A. Anghel, and K. Atasu, “Realistic synthetic financial transactions for anti-money laundering models,” inAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023

    2023

  16. [16]

    Faster homomorphic linear transformations in HElib,

    S. Halevi and V . Shoup, “Faster homomorphic linear transformations in HElib,” inAdvances in Cryptology – CRYPTO 2018, ser. Lecture Notes in Computer Science, vol. 10991. Springer, 2018, pp. 93–120

    2018

  17. [17]

    GEARSage: Graph edge-aware GraphSAGE for DGraphFin,

    J. Li, Z. Yu, W. Sun, X. Jin, Q. Wang, and L. Chen, “GEARSage: Graph edge-aware GraphSAGE for DGraphFin,” 7th Finvolution Data Science Competition, technical report, 2022, Sun Yat-sen University, SYSU-GEAR team. [Online]. Available: https://github.com/ storyandwine/GEARSage-DGraphFin

    2022

  18. [18]

    Microsoft SEAL (release 4.1),

    Microsoft Research, “Microsoft SEAL (release 4.1),” 2024, homomorphic encryption library. [Online]. Available: https: //github.com/microsoft/SEAL

    2024

  19. [19]

    Graph transformer networks,

    S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim, “Graph transformer networks,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019

    2019

  20. [20]

    SIGN: Scalable inception graph neural networks,

    F. Frasca, E. Rossi, D. Eynard, B. Chamberlain, M. Bronstein, and F. Monti, “SIGN: Scalable inception graph neural networks,”arXiv preprint arXiv:2004.11198, 2020

    arXiv 2004

  21. [21]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  22. [22]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li`o, and Y . Ben- gio, “Graph attention networks,” inProc. International Conference on Learning Representations (ICLR), 2018

    2018

  23. [23]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inProc. International Conference on Learning Representations (ICLR), 2017

    2017