Pith. sign in

REVIEW 4 major objections 7 minor 44 references

FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read FicGCN claims a homomorphic-encryption framework that runs privacy-preserving GCN inference up to 4.10x faster than prior designs, by exploiting the irregular sparsity of the graph's adjacency matrix.

desk verdict The 4.10x headline is internally inconsistent across tables; the underlying techniques are real but the numbers need reconciliation before the speedup claim can be trusted. read the letter →

arxiv 2506.10399 v1 pith:NHZQX3YR submitted 2025-06-12 cs.CR

classification cs.CR
keywords homomorphicencryptionCKKSgraphconvolutionalnetworksprivateinferencesparseaggregationciphertextpackingnodeorderoptimizationlatency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes FicGCN, a framework for running Graph Convolutional Network (GCN) inference on homomorphically encrypted data, so that a cloud server can compute on a client's graph features without ever seeing them. Its claim is that the irregular sparsity of the adjacency matrix is a resource, not a nuisance: by packing the encrypted feature matrix in a latency-aware way, gathering neighbours inside a single ciphertext with masked rotations, and reordering nodes around the ciphertext ring, the dominant cost of homomorphic rotations is sharply reduced. Across four benchmark datasets the paper reports the best end-to-end latencies, culminating in 7,733 seconds on the 19,793-node Corafull graph versus 31,735 seconds for the previous fastest design, a 4.10x improvement. If the numbers hold, privacy-preserving graph services for domains such as healthcare and finance move closer to practice, with encrypted inference on a 20,000-node graph shrinking to about two hours on a single CPU thread.

What carries the argument

The load-bearing mechanism is SpIntra-CA (Sparse Intra-Ciphertext Aggregation), an adaptation of the classic power-of-two ciphertext internal-sum trick to irregular graphs: each node's required rotation length is decomposed into bits, the ciphertext is rotated by $2^0, 2^1, \ldots$, and after each step a mask plaintext polynomial (a vector of 0s and 1s) keeps the slots that have reached their target neighbour and removes those that have finished. This replaces the one-rotation-per-node strategy with roughly $\log(N)$ rotations in favourable cases, at the cost of possible slot conflicts when two nodes want the same position; the paper's worst-case analysis gives $O(n\log^2 N)$ rotations and notes that the exact relationship between conflicts and ciphertext sparsity is hard to derive because it touches the NP-complete graph-colouring problem, so the practical gains are empirical rather than proven bounds. Two supporting devices carry the efficiency: a latency-aware packing scheme that chooses the number of columns $t$ per ciphertext by minimising an objective like $2\lceil Fn/t\rceil + 20\lceil\log(t)\rceil$, and Node Order Optimization (NOO), an offline reordering that groups sibling nodes into regions, interleaves regions around the ciphertext ring, and greedily minimises conflicts.

What would settle it

Re-run FicGCN and CryptoGCN end-to-end on Corafull on the same single-threaded machine using Microsoft SEAL 3.7.2 with the paper's stated parameters ($\Delta=2^{30}$, $M=2^{12}$, $Q=2^{18}$) and compare latencies: if FicGCN does not finish near 7,733 seconds while CryptoGCN costs near 31,735 seconds, the central 4.10x claim does not reproduce.

Watch

Extended reading notes

Core claim

The paper's central claim is that the irregular sparsity of a GCN's adjacency matrix can be converted directly into speed instead of being fought against. FicGCN packs the encrypted feature matrix column-wise with a per-dataset choice of how many columns share a ciphertext, then decides layer by layer whether to aggregate across ciphertexts (Inter-CA) or within a single ciphertext (SpIntra-CA). In SpIntra-CA, a single packed ciphertext is rotated by powers of two and a 0/1 mask is applied after each rotation, so that every slot simultaneously receives a neighbour or keeps its own value, reaching the target neighbour arrangement in roughly $\log(N)$ rotations in the ideal case. A node-order optimization pass partitions the graph into sibling-dense regions, interleaves those regions on the ciphertext ring, and greedily places nodes to avoid slot conflicts, which keeps the extra rotations caused by collisions low. With this combination, the paper reports the fastest inference on Cora, Citeseer, Corafull and NTU-cross-View, with the largest gain, 4.10x over CryptoGCN, on Corafull, the biggest graph tested.

Load-bearing premise

The speedup claims rest on the presumption that the Gazelle, Penguin, and CryptoGCN baseline latencies in Table 3 were measured on the same hardware and with the same ciphertext parameters as FicGCN; the paper states its environment (single-threaded Intel i7-9750H, SEAL 3.7.2) but releases no code, so the baseline figures cannot be independently replayed.

Editorial extensions

If this is right

  • On the paper's reported numbers, encrypted inference on the 19,793-node Corafull graph drops from 31,735 seconds (CryptoGCN) to 7,733 seconds (FicGCN), a 4.10x gain, and the gap to the oldest baseline, Gazelle, exceeds 120x.
  • The gains concentrate where the graph is large relative to the ciphertext slot count: the speedup over the fastest baseline grows from 1.26x on the 25-node NTU dataset to 2.01x on Cora and 4.10x on Corafull.
  • Ablation results attribute most of the benefit to rotation reduction: Node Order Optimization alone cuts rotations by about 66% on Cora and Citeseer relative to un-reordered SpIntra-CA, and the latency-aware packing choice is worth about 1.27x-1.44x on those datasets.
  • The offline node-reordering overhead stays comparatively tiny even at scale: on a million-node Pokec estimate the paper measures preprocessing at about 0.002% of the online encrypted-inference time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's cost model — a per-layer choice between sparse intra-ciphertext rotation and dense inter-ciphertext multiplication, driven by feature dimension versus slot count — is a template that could transfer to other CKKS workloads that multiply a sparse plaintext matrix against packed ciphertexts, such as encrypted recommendation or attention layers.
  • Because the node order inside a ciphertext is a free variable, NOO's region-interleaving idea suggests a testable extension for other message-passing models: re-run the region detection against a given neighbour-sampling scheme and check whether rotation counts fall in proportion to how tightly sibling nodes are co-located on the ring.
  • The paper's own caveat that conflict behaviour connects to the NP-complete graph-colouring problem implies that the reported speedups are empirical, heuristic results rather than proven bounds, so the method's edge over dense baselines should be re-checked whenever the graph's degree distribution or the sampler changes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes FicGCN, a CKKS-based framework for privacy-preserving GCN inference. The framework consists of three main techniques: (i) a latency-aware column-based packing strategy, (ii) a sparse intra-ciphertext aggregation method (SpIntra-CA) with aggregation order and ciphertext processing order optimizations (AOO and CPOO), and (iii) a node order optimization (NOO) algorithm based on BFS region partitioning, interleaved region arrangement, and a greedy conflict-minimizing search. The authors evaluate FicGCN on Cora, Citeseer, Corafull, and NTU datasets against Gazelle, Penguin, and CryptoGCN, reporting speedups up to 4.10x over CryptoGCN on Corafull, and claim the best performance across all tested datasets.

Significance. If the reported speedups (up to 4.10x) hold under a fair comparison, this is a notable practical advance for HE-based GCN inference: the paper directly targets the rotation bottleneck created by irregular graph sparsity and combines it with a realizable packing and reordering pipeline. The paper's strengths include a clear operational breakdown (Rot/PMult/Add), a worst-case analysis of the SpIntra-CA rotation overhead, ablation studies for AOO, CPOO, and NOO, and an overhead analysis of NOO on a large-scale graph (Pokec). However, the headline result currently rests on internally inconsistent measurements across the main text and appendix, and the fairness of the baseline comparison is not fully documented. These issues must be resolved before the empirical claim can be accepted.

major comments (4)
  1. [Tables 3, 4, and Appendix Table 8] The latency reported for the same 'w/ NOO' configuration is inconsistent across tables: for Corafull, Table 3 and Appendix Table 8 report 7,733s, while main-text Table 4 reports 9,075s (with Rot/PMult/Add counts of 36.7K/14.7M/16.9M vs. 32.1K/12.7M/14.9M). Similar discrepancies appear for Cora (64.12s vs. 69.28s), Citeseer (79.98s vs. 86.14s), and NTU (1373.82s vs. 1463.80s). If Table 4 is correct, the claimed 4.10x speedup on Corafull would become 31,735/9,075 = 3.50x, which contradicts the abstract and conclusion. The authors must identify which table is correct and explain the discrepancy.
  2. [Sections 4.1 and 4.2.1] The manuscript does not explicitly state whether the Gazelle, Penguin, and CryptoGCN baseline latencies in Table 3 were re-implemented and measured locally on the same Intel i7-9750H / SEAL 3.7.2 platform. If these numbers were taken from prior publications with different hardware or HE parameter settings, the claimed speedups would not be fair comparisons. Please clarify the provenance of each baseline number and, if they were reproduced, describe the re-implementation and validation steps.
  3. [Section 2.5 vs. Section 4.1 / Appendix Table 6] Section 2.5 states that all HE parameters are selected to achieve 128-bit security, but Appendix Table 6 lists the NTU configuration with security level '>=80-bit'. Since the paper's privacy guarantee rests on this claim, the contradiction must be resolved, either by adjusting the security statement or by providing parameter estimates that justify 128-bit for the NTU settings.
  4. [Section 3.3.2 and Appendix B] The mode-selection formula for SpIntra-CA in Section 3.3.2 uses the term 10cn log2(N) with c in [0,1] a free parameter, while the cost model in Appendix B.1 assumes Rot is 20x slower than PMult/Add. The factor 10 and the choice of c are not derived, and no sensitivity analysis is provided. Because the abstract claims a 'globally optimal balance' between aggregation and combination, this heuristic should be justified or the claims should be softened to describe a heuristic scheduling strategy.
minor comments (7)
  1. [Section 2.3] The name 'Peinguin' should be 'Penguin'.
  2. [Section 4.2.1 and Table 3] The text claims '>120x over Gazelle' on Corafull, but Table 3 lists Gazelle's Corafull latency as '/'; explain how this speedup bound was obtained or remove it.
  3. [Section 4.3] The text says 'The results are presented in Table 4', but the overhead analysis appears in Table 5.
  4. [Appendix A] The phrase 'the delay caused by Rot exceeds A%' uses an undefined symbol 'A%'; this should be a concrete percentage or the placeholder should be removed.
  5. [Section 4.1] The notation 'Q=2^18' and 'Q=680' is nonstandard for a CKKS modulus; clarify whether these are bit lengths or actual modulus sizes, and use consistent notation throughout.
  6. [Appendix C.2 and C.3] The cross-references to appendix tables are inconsistent: Section C.2 refers to 'Table 6' when the packing ablation is Table 7, and Section C.3 refers to 'Table 7' when the NOO ablation is Table 8.
  7. [Algorithm 1] The variable 'n' is used both as a node variable and as the sampled neighbor count; rename one of them to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the 4.10x speedup claim is an end-to-end measured latency result against external baselines; the cost-model constants (c, '10', '20') are design heuristics that do not force the measured numbers.

full rationale

FicGCN's central claim — an end-to-end speedup of up to 4.10x versus CryptoGCN on Corafull (Table 3: 7,733s vs 31,735s) — is a measured latency result, not a quantity derived from the paper's own fitted constants. The two cost models — the packing objective J(t;F,n) = 2⌈F·n/t⌉ + 20⌈log(t)⌉ in Section 3.2/Appendix B and the aggregation-mode selector Agg = arg min{Inter-CA: 2⌈F·n/t⌉, SpIntra-CA: 10cnlog2(N)} in Section 3.3.2 — are heuristics that choose a packing parameter t and a per-layer aggregation mode; the reported latencies are then measured empirically (Ablation Tables 1, 4, 7, 8), so the headline speedup does not reduce by construction to the constants '20', '10', or to the ad hoc factor c∈[0,1] introduced 'to quantify the effect of the optimization (like AOO) and the actual Rot number in practice.' The SOTA baselines (Gazelle, Penguin, CryptoGCN) come from external author groups (Juvekar et al.; Ran et al. 2022/2024, with no author overlap with the present paper), so no self-citation chain is load-bearing, and no uniqueness theorem is imported. The derivation chain is self-contained against external benchmarks, and the circularity burden is low. Non-circularity concerns that should nevertheless be flagged for the correctness pass: (1) the paper never states that the Table 3 baselines were re-implemented and run locally on the same single-thread i7-9750H / Microsoft SEAL 3.7.2 setup, and no code is released, so the fairness of the 1.26x-4.10x ratios rests on an unstated assumption; (2) the paper's own tables disagree on the headline FicGCN+NOO numbers — Corafull is 7,733s in Table 3 and Table 8 but 9,075s in Table 4 (Cora 64.12s vs 69.28s; Citeseer 79.98s vs 86.14s; NTU 1,373.82s vs 1,463.80s), which would change the advertised 4.10x to about 3.50x if Table 4 reports the true measurements; and (3) Appendix C.2's conclusion that the optimal t minimizes first-layer Inter-CA overhead is in tension with its own Cora data (t=1: 46.02s vs t=2: 36.23s overall). These are data-integrity and measurement-validity issues, not definitional circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central speedup claim rests on a small number of hand-tuned parameters (c, TH, n) and on cost-model assumptions (fixed Rot/PMult ratio, offline first layer). These are not fatal, but they mean the globally optimal framing is heuristic rather than a derived optimum.

free parameters (4)
  • c (SpIntra-CA conflict factor) = not reported (stated as c in [0,1])
    Introduced in Section 3.3.2 to scale the worst-case rotation count n*log2(N) to practical values; the paper does not state how c is determined or its value, making the mode-selection formula untestable.
  • Mode-selection constant 10 = 10
    The formula Agg=argmin{2*ceil(Fn/t), 10cn log2(N)} uses a factor 10, while the text states rotations are over 20x slower than PMult and Add; no justification for this discrepancy is given.
  • Region size threshold TH = 1024 (Cora, Citeseer), 4096 (Corafull)
    Chosen by hand to control NOO region size; affects conflict reduction but not correctness.
  • Sampled neighbor count n = approximately average degree (4.01, 2.85, 12.82, etc.)
    The number of neighbors sampled in GraphSage is set to approximately the average degree per dataset, a configuration choice that affects the Inter-CA versus SpIntra-CA trade-off.
assumptions (4)
  • domain assumption Rotations are consistently more than 20x slower than plaintext-ciphertext multiplications and additions across the relevant modulus levels.
    Used in Appendix A and Appendix B to derive the objective function J(t) and the mode-selection rule; the actual ratio is implementation- and level-dependent and is not measured in the paper.
  • domain assumption The client can perform the first-layer Inter-CA encryptions offline and at negligible cost, so only server-side latency is counted.
    Section 3.3.2 states that the multiple inter-ciphertext computations can be eliminated by offline user encryptions from the client, which can be ignored; this excludes client encryption time from all reported latencies.
  • standard math The CKKS parameters chosen (ring degree 4096 or 8192, modulus bit-length 218 or 680) provide the stated 128-bit security.
    Section 2.5 asserts 128-bit security by citing Cheon et al. 2018 but provides no concrete security estimation; this is a standard assumption in HE systems papers.
  • domain assumption Each node has exactly n sampled neighbors after GraphSage sampling, so the number of neighbor ciphertexts is fixed at n.
    The complexity analysis in Section 3.3 and Table 2 assumes n per-node neighbor ciphertexts; nodes with fewer than n neighbors would require padding or sampling with replacement, which is not discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks." pith.science (2026). https://pith.science/paper/NHZQX3YR

@misc{pith2026250610399,
  author       = {Pith},
  title        = {Pith review of: FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHZQX3YR}},
  note         = {Machine review of arXiv:2506.10399}
}
read the original abstract

Graph Convolutional Neural Networks (GCNs) have gained widespread popularity in various fields like personal healthcare and financial systems, due to their remarkable performance. Despite the growing demand for cloud-based GCN services, privacy concerns over sensitive graph data remain significant. Homomorphic Encryption (HE) facilitates Privacy-Preserving Machine Learning (PPML) by allowing computations to be performed on encrypted data. However, HE introduces substantial computational overhead, particularly for GCN operations that require rotations and multiplications in matrix products. The sparsity of GCNs offers significant performance potential, but their irregularity introduces additional operations that reduce practical gains. In this paper, we propose FicGCN, a HE-based framework specifically designed to harness the sparse characteristics of GCNs and strike a globally optimal balance between aggregation and combination operations. FicGCN employs a latency-aware packing scheme, a Sparse Intra-Ciphertext Aggregation (SpIntra-CA) method to minimize rotation overhead, and a region-based data reordering driven by local adjacency structure. We evaluated FicGCN on several popular datasets, and the results show that FicGCN achieved the best performance across all tested datasets, with up to a 4.10x improvement over the latest design.

Figures

Figures reproduced from arXiv: 2506.10399 by the authors.

Figure 1
Figure 1. The contradiction between the computing characteristics of GCN and CKKS. multiplication AXW where A and W are plaintexts and X is ciphertext, which includes a large number of rotations and multiplications. Drawing inspiration from the efficiency optimization of plaintext GCN computations, recent work leverages the sparsity of A for optimization. To this end, an adjacency-matrix-aware data packing and multiplication … view at source ↗
Figure 2
Figure 2. Workflow of FicGCN. (a) Three main steps of the HE-specific reordering algorithm: Detect, Arrange and Search. (b) The computational complexity comparison between Inter-CA and SpIntra-CA, SpIntra-CA encompasses log(N) fundamental rotations along with several additional rotations triggered by conflicts. (The dashed lines represent optional data paths, and the modules shown in detail are those that represent the optimi… view at source ↗
Figure 3
Figure 3. A toy example of SpIntra-CA. (a) Aggregation in prior work node by node. (b) The process of SpIntra-CA. (c) The NOO effect. (d) Conflicts arise when multiple nodes occupy the same slot, and removals occur when a node reaches its aggregation target position. Also CPOO may help merge some common Rots in sparse ciphertexts (b)The relationship between slot utilization and the index of additional ciphertexts in SpIntra-C… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Aggregation and ciphertext processing order problem in SpIntra-CA and removal in SpIntra-CA lead to more sparse ciphertext, which decreases the efficiency of SIMD. Observing that no conflicts occur in continuous multiple rotations on sparse ciphertext, we propose to me…
Figure 5
Figure 5. Figure 5: Node order optimization (NOO). Interleaved region ar￾rangement (b) requires fewer rotations than the region-by-region arrangement (a). Greedy node arrangement for least conflicts within each region (c) we enqueued sibling nodes instead of neighboring nodes in each iter…
Figure 6
Figure 6. Figure 6: The effectiveness of CPOO. (a) The relationship between latency and utilization thresholds across four datasets. (b) The ratio of Rots pruned by CPOO of 4 datasets. (c) Stage-wise ciphertext number w/ and w/o CPOO [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: (a) illustrates the HE operation individual cost at different levels for ciphertexts under the HE parameter configuration used for large graph inference in this paper. It shows that the overhead of Rot and CMult is often over 20× greater than that of PMult and Add [PI…
Figure 8
Figure 8. Figure 8: (a) An example objective function (b) The principal of node order back propagation which influences the initial packing order from the client. C. Supplementary Experiments C.1. Experiments Setup All experimental settings are mentioned in Section 4.1. Here, this informa…
Figure 9
Figure 9. Figure 9: Interleave Arrangement Effect of 4 Datasets Algorithm 1 Node Order Optimization for SpIntra-CA Input: Graph: G = (V, E) ; M: Slot number; T H: Maximum number of nodes in one region; Ae: Adjacency matrix after neighbor￾sampling; BF Ssibling: A modified breadth-first sea…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 36 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Boddeti, V

    Ao, W. and Boddeti, V. N. (2024). \ AutoFHE \ : Automated adaption of \ CNNs \ for efficient evaluation over \ FHE \ . In 33rd USENIX Security Symposium (USENIX Security 24) , pages 2173--2190

  3. [3]

    Arai, J., Shiokawa, H., Yamamuro, T., Onizuka, M., and Iwamura, S. (2016). Rabbit order: Just-in-time parallel reordering for fast graph analysis. In 2016 IEEE International Parallel and Distributed Processing Symposium (IPDPS) , pages 22--31. IEEE

  4. [4]

    H., Hanrot, G., Park, J

    Bae, Y., Cheon, J. H., Hanrot, G., Park, J. H., and Stehl \'e , D. (2024). Plaintext-ciphertext matrix multiplication and fhe bootstrapping: Fast and fused. In Annual International Cryptology Conference , pages 387--421. Springer

  5. [5]

    Bongini, P., Bianchini, M., and Scarselli, F. (2021). Molecular generative graph neural networks for drug discovery. Neurocomputing , 450:242--252

  6. [6]

    Brakerski, Z., Gentry, C., and Vaikuntanathan, V. (2014). (leveled) fully homomorphic encryption without bootstrapping. ACM Transactions on Computation Theory (TOCT) , 6(3):1--36

  7. [7]

    Brutzkus, A., Gilad-Bachrach, R., and Elisha, O. (2019). Low latency privacy preserving inference. In International Conference on Machine Learning , pages 812--821. PMLR

  8. [8]

    Chen, T., Bao, H., Huang, S., Dong, L., Jiao, B., Jiang, D., Zhou, H., Li, J., and Wei, F. (2022). The-x: Privacy-preserving transformer inference with homomorphic encryption. arXiv preprint arXiv:2206.00216

Show all 44 references
  1. [9]

    H., Han, K., Kim, A., Kim, M., and Song, Y

    Cheon, J. H., Han, K., Kim, A., Kim, M., and Song, Y. (2018). Bootstrapping for approximate homomorphic encryption. In Advances in Cryptology--EUROCRYPT 2018: 37th Annual International Conference on the Theory and Applications of Cryptographic Techniques, Tel Aviv, Israel, Apr...

  2. [10]

    H., Kim, A., Kim, M., and Song, Y

    Cheon, J. H., Kim, A., Kim, M., and Song, Y. (2017). Homomorphic encryption for arithmetic of approximate numbers. In Advances in Cryptology--ASIACRYPT 2017: 23rd International Conference on the Theory and Applications of Cryptology and Information Security, Hong Kong, China, ...

  3. [11]

    S., and Kim, H

    Choi, H., Woo, S. S., and Kim, H. (2024). Blind-touch: Homomorphic encryption-based distributed neural network inference for privacy-preserving fingerprint authentication. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 21976--21985

  4. [12]

    Geng, T., Wu, C., Zhang, Y., Tan, C., Xie, C., You, H., Herbordt, M., Lin, Y., and Li, A. (2021). I-gcn: A graph convolutional network accelerator with runtime locality enhancement through islandization. In MICRO-54: 54th annual IEEE/ACM international symposium on microarchite...

  5. [13]

    Gilad-Bachrach, R., Dowlin, N., Laine, K., Lauter, K., Naehrig, M., and Wernsing, J. (2016). Cryptonets: Applying neural networks to encrypted data with high throughput and accuracy. In International conference on machine learning , pages 201--210. PMLR

  6. [14]

    and Shoup, V

    Halevi, S. and Shoup, V. (2014). Algorithms in helib. In Advances in Cryptology--CRYPTO 2014: 34th Annual Cryptology Conference, Santa Barbara, CA, USA, August 17-21, 2014, Proceedings, Part I 34 , pages 554--571. Springer

  7. [15]

    Hamilton, W., Ying, Z., and Leskovec, J. (2017). Inductive representation learning on large graphs. Advances in neural information processing systems , 30

  8. [16]

    Hao, M., Li, H., Chen, H., Xing, P., Xu, G., and Zhang, T. (2022). Iron: Private inference on transformers. Advances in neural information processing systems , 35:15718--15731

  9. [17]

    Huang, H. (2024). Investigation of CKKS-Induced Error on Accuracy Changes of GraphSAGE Inference with Fully Homomorphic Encryption . PhD thesis, Waseda University

  10. [18]

    K., Ghodsi, Z., Garg, S., and Reagen, B

    Jha, N. K., Ghodsi, Z., Garg, S., and Reagen, B. (2021). Deepreduce: Relu reduction for fast private inference. In International Conference on Machine Learning , pages 4839--4849. PMLR

  11. [19]

    Jia, Z., Lin, S., Ying, R., You, J., Leskovec, J., and Aiken, A. (2020). Redundancy-free computation for graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages 997--1005

  12. [20]

    Jin, W., Yao, Y., Han, S., Gu, J., Joe-Wong, C., Ravi, S., Avestimehr, S., and He, C. (2023). Fedml-he: An efficient homomorphic-encryption-based privacy-preserving federated learning system. arXiv preprint arXiv:2303.10837

  13. [21]

    H., and Ahn, J

    Jung, W., Lee, E., Kim, S., Kim, J., Kim, N., Lee, K., Min, C., Cheon, J. H., and Ahn, J. H. (2021). Accelerating fully homomorphic encryption through architecture-centric analysis and optimization. IEEE Access , 9:98772--98789

  14. [22]

    Juvekar, C., Vaikuntanathan, V., and Chandrakasan, A. (2018). \ GAZELLE \ : A low latency framework for secure neural network inference. In 27th USENIX security symposium (USENIX security 18) , pages 1651--1669

  15. [23]

    Kipf, T. N. and Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

  16. [24]

    Lee, E., Lee, J.-W., Lee, J., Kim, Y.-S., Kim, Y., No, J.-S., and Choi, W. (2022). Low-complexity deep convolutional neural networks on fully homomorphic encryption using multiplexed parallel convolutions. In International Conference on Machine Learning , pages 12403--12422. PMLR

  17. [25]

    W., Shin, J., and Lee, M.-K

    Lee, S., Lee, G., Kim, J. W., Shin, J., and Lee, M.-K. (2023). Hetal: efficient privacy-preserving transfer learning with homomorphic encryption. In International Conference on Machine Learning , pages 19010--19035. PMLR

  18. [26]

    and Jiang, L

    Lou, Q. and Jiang, L. (2021). Hemet: A homomorphic-encryption-friendly privacy-preserving mobile neural network architecture. In International conference on machine learning , pages 7102--7110. PMLR

  19. [27]

    Matsunaga, D., Suzumura, T., and Takahashi, T. (2019). Exploring graph neural networks for stock market predictions with rolling window analysis. arXiv preprint arXiv:1909.10660

  20. [28]

    Merkel, N., Toussing, P., Mayer, R., and Jacobsen, H.-A. (2024). Can graph reordering speed up graph neural network training? an experimental study. arXiv preprint arXiv:2409.11129

  21. [29]

    Peng, H., Ran, R., Luo, Y., Zhao, J., Huang, S., Thorat, K., Geng, T., Wang, C., Xu, X., Wen, W., et al. (2024). Lingcn: Structural linearized graph convolutional network for homomorphically encrypted inference. Advances in Neural Information Processing Systems , 36

  22. [30]

    Ran, R., Luo, X., Wang, W., Liu, T., Quan, G., Xu, X., Ding, C., and Wen, W. (2023). Spencnn: orchestrating encoding and sparsity for fast homomorphically encrypted neural network inference. In International Conference on Machine Learning , pages 28718--28728. PMLR

  23. [31]

    Ran, R., Wang, W., Gang, Q., Yin, J., Xu, N., and Wen, W. (2022). Cryptogcn: Fast and scalable homomorphically encrypted graph convolutional network inference. Advances in Neural information processing systems , 35:37676--37689

  24. [32]

    Ran, R., Xu, N., Liu, T., Wang, W., Quan, G., and Wen, W. (2024). Penguin: parallel-packed homomorphic encryption for fast graph convolutional network inference. Advances in Neural Information Processing Systems , 36

  25. [33]

    T., Lee, H.-H

    Reagen, B., Choi, W.-S., Ko, Y., Lee, V. T., Lee, H.-H. S., Wei, G.-Y., and Brooks, D. (2021). Cheetah: Optimizing and accelerating homomorphic encryption for private inference. In 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA) , pages 26--39. IEEE

  26. [34]

    Ru, S., Zhang, B., Jie, Y., Zhang, C., Wei, L., and Gu, C. (2021). Graph neural networks for privacy-preserving recommendation with secure hardware. In 2021 international conference on networking and network applications (NaNA) , pages 395--400. IEEE

  27. [35]

    Si, C., Jing, Y., Wang, W., Wang, L., and Tan, T. (2018). Skeleton-based action recognition with spatial reasoning and temporal stack learning. In Proceedings of the European conference on computer vision (ECCV) , pages 103--118

  28. [36]

    Z., Akshayaram, P., and Ada, P

    Srinivasan, W. Z., Akshayaram, P., and Ada, P. R. (2019). Delphi: A cryptographic inference service for neural networks. In Proc. 29th USENIX secur. symp , volume 3

  29. [37]

    X., Lu, C., and Lin, X

    Wei, H., Yu, J. X., Lu, C., and Lin, X. (2016). Speedup graph processing by graph ordering. In Proceedings of the 2016 International Conference on Management of Data , pages 1813--1828

  30. [38]

    Wu, H., Fang, W., Zheng, Y., Ma, J., Tan, J., Wang, Y., and Wang, L. (2024). Ditto: Quantization-aware secure inference of transformers upon mpc. arXiv preprint arXiv:2405.05525

  31. [39]

    Wu, S., Sun, F., Zhang, W., Xie, X., and Cui, B. (2022). Graph neural networks in recommender systems: a survey. ACM Computing Surveys , 55(5):1--37

  32. [40]

    Yan, S., Xiong, Y., and Lin, D. (2018). Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence , volume 32

  33. [41]

    Zeng, W., Li, M., Xiong, W., Tong, T., Lu, W.-j., Tan, J., Wang, R., and Huang, R. (2023a). Mpcvit: Searching for accurate and efficient mpc-friendly vision transformer with heterogeneous attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision , p...

  34. [42]

    Zeng, W., Li, M., Yang, H., Lu, W.-j., Wang, R., and Huang, R. (2023b). Copriv: network/protocol co-optimization for communication-efficient private inference. Advances in Neural Information Processing Systems , 36:78906--78925

  35. [43]

    Zhang, C., Li, S., Xia, J., Wang, W., Yan, F., and Liu, Y. (2020). \ BatchCrypt \ : Efficient homomorphic encryption for \ Cross-Silo \ federated learning. In 2020 USENIX annual technical conference (USENIX ATC 20) , pages 493--506

  36. [44]

    Zhu, L., Hua, Q.-s., Chen, Y., and Jin, H. (2023). Secure outsourced matrix multiplication with fully homomorphic encryption. In European Symposium on Research in Computer Security , pages 249--269. Springer

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.