Pith. sign in

REVIEW 4 major objections 6 minor 42 references

Technical Evaluation of a Disruptive Approach in Homomorphic AI

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A keyed hash that preserves similarity lets off-the-shelf AI algorithms work directly on encrypted data, this independent evaluation confirms.

desk verdict Self-evaluation of a proprietary scheme whose own MLP result refutes the central 'no information loss' claim; useful as a cautionary case but not a scientific contribution. read the letter →

arxiv 2506.11954 v1 pith:ZR6ECMPV submitted 2025-06-13 cs.CR cs.AI

Eric Filiol This is my paper
classification cs.CRcs.AI
keywords homomorphicencryptionhash-basedAIkeyedhashfunctionsimilarity-preservingprivacy-preservingmachinelearningclusteringonencrypteddataFashion-MNISTnon-malleability
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 evaluates HbHAI, a proprietary scheme that encrypts data with a keyed compression hash while preserving the similarity structure that AI algorithms rely on. The author ran unmodified clustering, classification, and neural-network code on two HbHAI-protected datasets and found that clustering partitions were identical to those on plaintext (Rand index 1), k-NN classifications matched exactly, and a two-step random forest on protected Fashion-MNIST reached 95.34% validation accuracy at compression rate $\delta=3$ and 93.8% at $\delta=6$. Data size shrank by about $3\times$ and computation time by $2.8\times$ with a dedicated C implementation; off-the-shelf libraries showed only about 18% time savings due to their own overhead, still negligible compared with the million-fold slowdown of fully homomorphic encryption. The paper concludes that no information loss occurs and that models are preserved, while cautioning that security claims rest on a black-box analysis and need white-box verification.

What carries the argument

The load-bearing object is the HAI hash function class $H_{K,\delta}$ of Definition 1: a keyed, compressing function whose defining property is similarity preservation — for any three objects, the ordering of similarities in the plaintext domain is exactly the ordering in the hashed domain. This equivalence is what lets any similarity-based AI algorithm consume protected data as if it were plaintext. A companion definition (Definition 2) requires the hash to be pre-image resistant, collision- and second-preimage-resistant, and non-malleable, so that inputs cannot be recovered and models cannot be extracted. The evaluation machinery that carries the empirical argument is the Rand index for comparing clustering partitions, $k$-modes and $k$-NN implementations in C using a multi-precision arithmetic library, and standard deep-learning and decision-tree libraries for the Fashion-MNIST benchmark.

What would settle it

Take a sample of the protected Fashion-MNIST images at $\delta=3$, compute nearest-neighbor rankings for a set of images using a similarity measure the paper did not test, such as cosine similarity, and compare those rankings against the plaintext images; if any triple reverses its similarity ordering, the central preservation claim fails. A complementary test is a key-recovery or linkage attempt on the protected cyber dataset, exploiting the known class labels and the keyed intra-class permutation, to see whether protected records can be matched to plaintext records.

Watch

Extended reading notes

Core claim

HbHAI's central claim is that a keyed hash function $H_{K,\delta}$ with compression rate $\delta$ can simultaneously compress data and preserve similarity order: for a given similarity measure $S$ and any three objects $x,x',x''$, the inequality $S(x,x'')<S(x,x')$ holds if and only if $S(H_{K,\delta}(x),H_{K,\delta}(x''))<S(H_{K,\delta}(x),H_{K,\delta}(x'))$. The evaluation tests this claim on a cybersecurity dataset of 2,200 objects with 49,955 categorical features and on Fashion-MNIST. With a custom C implementation of $k$-modes clustering and $k$-NN on the protected cyber data, clustering achieved a Rand index of exactly 1 against plaintext and classification results were identical. On Fashion-MNIST, an unmodified three-layer MLP reached 81.88% validation accuracy at $\delta=3$ versus 88.54% on plaintext, but a two-stage random forest built to handle a poorly predicted class reached 95.34%, the sixth-best of 39 published results; at $\delta=6$ a two-model approach reached 93.8%. The paper reports an observed data-size reduction of $2.99\times$ and a $2.8\times$ speed-up with the dedicated implementation, and at the black-box level no way to link protected records to plaintext, so it confirms most of HbHAI's claims with only minor reservations.

Load-bearing premise

The load-bearing premise is that a keyed hash family exists which simultaneously delivers the cryptographic properties of Definition 2 (pre-image resistance, collision resistance, model security, non-malleability) and the strict similarity-order preservation of Definition 1; the paper takes this existence from its companion formalization [12] without proof, and its evaluation sees only a black-box implementation.

Editorial extensions

If this is right

  • If HbHAI is secure, privacy-preserving cloud AI can run on existing, unmodified machine-learning libraries, eliminating the complete algorithm rewrites that current fully homomorphic encryption systems require.
  • Compression by a factor $\delta$ between 3 and 6 means protected data needs a fraction of the storage and bandwidth of plaintext, and with optimized implementations processing time drops by a comparable factor.
  • The reduction in storage and compute directly lowers the energy footprint of outsourced AI and makes encrypted processing feasible on constrained hardware such as single-board computers and embedded systems.
  • Because results are mapped back to plaintext through object indices rather than through decryption, the workflow requires only that the data owner retains the index correspondence, not that the processing party ever sees plaintext.
  • The recovery of 95% accuracy on protected Fashion-MNIST with a two-stage random forest shows that any accuracy loss introduced by the hash can be offset with model-level adjustments, independent of the cryptographic layer.

Reading between the lines

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

  • If similarity-order preservation is a genuine property of the hash, then the scheme should extend beyond clustering and classification to any ranking-based AI task, such as nearest-neighbor search, information retrieval, and recommendation; this generalization goes beyond what the two datasets demonstrate.
  • The performance comparison is drawn against fully homomorphic encryption's million-fold slowdown, but not against other privacy technologies such as secure enclaves or federated learning, so the practical cost picture depends on the baseline one chooses.
  • Because the data owner must retain the plaintext-to-protected index mapping to transpose results, 'encryption' here acts more like a keyed, lossy transform with a trusted owner than like public-key encryption; its security model should be evaluated on those terms.
  • A decisive untested case is feeding the protected Fashion-MNIST images through a convolutional neural network, since the paper's fully connected and tree models cannot distinguish preserved metric structure from mere class separability.
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 / 6 minor

Summary. The manuscript reports a technical evaluation of HbHAI, a proprietary keyed-hash homomorphic AI scheme developed by Hope4Sec. The paper claims that HbHAI-protected data can be processed by unmodified, off-the-shelf AI algorithms with no information loss while reducing data size and computing time. It presents two use cases: a two-class cybersecurity dataset evaluated with k-modes clustering and k-NN classification, and Fashion-MNIST evaluated with an unmodified MLP, custom random-forest pipelines, and comparisons to 38 published plaintext results. The paper concludes that HbHAI's claims are confirmed, with only minor reservations.

Significance. If substantiated, an order-preserving keyed hash with cryptographic security would be a major advance: it would eliminate the rewrite requirement and massive overhead of FHE. The paper's positive aspects include the explicit reporting of the unmodified MLP result, the use of standard metrics (Rand index, validation accuracy), and the description of the constrained hardware environment. Its weakness is that the evidence does not support the headline claims: the single unmodified-algorithm comparison shows a substantial accuracy drop, the high-accuracy results come from bespoke fitted pipelines without plaintext baselines, the security analysis is black-box only, and the scheme's formal foundations are cited to the author's own unpublished work. As presented, the paper is not a reliable basis for the claimed conclusions.

major comments (4)
  1. [§4.2, Test with TensorFlow/Keras (δ=3) - MLP Approach] The unmodified Géron three-layer MLP obtains validation accuracy 0.8188 on HbHAI-protected Fashion-MNIST versus 0.8854 on plaintext, a gap of 6.66 percentage points. This directly contradicts the abstract's claim that native AI algorithms can be used without modification with no information loss, and it is inconsistent with the conclusion that 'no information loss has occurred and models are very well preserved.' The subsequent recovery to 0.9534 uses a custom two-stage random forest with an empirically chosen decision threshold, which is not an unmodified off-the-shelf algorithm and is never run on the plaintext data. The comparison with 38 published plaintext results is therefore not a comparison of preservation, but of a bespoke pipeline against unrelated published baselines.
  2. [§4.1, Use-Case 1: Clustering/Classification of Cybersecurity Data] The section states that 'Only the HbHAI-protected version of the dataset was available to us,' yet two paragraphs later it claims that 'This very classic clustering program was run on plaintext data and on HbHAI-protected data' and reports Rand index R=1 and identical k-NN results. These statements are mutually contradictory. If the plaintext version was indeed unavailable, the central claim of perfect preservation for the cyber dataset is unsupported; if it was available, the earlier sentence must be corrected. This inconsistency is load-bearing because the cyber dataset is the only evidence for exact clustering preservation.
  3. [§3.2, Definitions 1 and 2] The entire scheme depends on the existence of a keyed hash family H_{K,δ} that simultaneously satisfies strict similarity-order preservation (Definition 1.3) and the cryptographic security properties (Definition 2). The manuscript only cites the author's own formalization [12] and provides no construction, no proof of existence, and no reduction to a standard hard problem. The implementation is proprietary and the security evaluation is limited to a black-box search for associations, which the paper itself concedes cannot provide a rigorous assessment. Without a public specification or proof, both the security and the 'no information loss' claims are not verifiable, and the negative MLP result gives reason to doubt the similarity-preservation property at the level claimed.
  4. [Abstract and §1] The abstract and Section 1 call the analysis 'independent' and 'blind,' but Section 1 states that the author 'launched a collaborative project' with Hope4Sec and published the formalization of HbHAI in [12], and the Acknowledgments thank Hope4Sec's Jaagup Sepp for sharing datasets and technical data. The manuscript is therefore not an independent third-party evaluation in the usual sense. This framing should be corrected, as it materially affects the weight a reader can place on the confirmatory conclusions.
minor comments (6)
  1. [§2 heading] The heading of Section 2 contains a typo: 'Homomorphic Encrypption Landscape' should be 'Homomorphic Encryption Landscape'.
  2. [§4.2 and Abstract] In Section 4.2, 'we ten considered' and 'A cecond model' are typographical errors; 'As a results' in the Abstract should be 'As a result'.
  3. [§4.2, Figures] The text refers to 'Figure 4.2' and 'Fig. 1' for the learning curves, but no figure appears in the manuscript; either include the figure or remove the reference.
  4. [Table 1 and §4.1] Table 1 gives the δ=6 protected size as 5.2 without a unit, and the relationship between the 49,955 features, the 2,620-byte files, and the '49,955-bit integer' representation in Section 4.1 is not explained.
  5. [§3.2, Definition 1] Definition 1 uses a compression rate δ that is stated to lie in [3,6]⊂R and appears in the denominator n/δ, but the reported experiments use δ=3 and δ=6; the intended domain and rounding rules should be stated precisely.
  6. [References [12] and [38]] References [12] and [38] are listed as 'to be published'; the evaluation therefore relies on material that is not yet accessible to referees or the general research community.

Circularity Check

1 steps flagged · score 6.0 of 10

Fashion-MNIST 'preservation' result is a post-hoc fitted accuracy presented as an independent confirmation, while the paper's own unmodified MLP shows a 6.66-point accuracy drop.

  1. fitted input called prediction [Section 4.2, 'Analysis (δ=3) - 2-step Random Forest Approach' and 'Results Comparison']
    "Step 2.- Than we used a second Random Forest Model (1,500 decision trees) to specifically process class 6 by using of an empiric decision threshold below which data are decided in class 6. As a result, we got a final validation accuracy of0.9534."

    The reported 95.34% validation accuracy is not an independent, out-of-sample measurement of HbHAI's preservation of plaintext models. The second-stage decision threshold is called 'empiric' and was introduced specifically because class 6 was 'generally badly predicted'; it is tuned on the same validation set whose accuracy is then reported as the result. The reported metric is thus the objective used to select the threshold, so the 'final validation accuracy' is statistically forced by the fit rather than predicted. Furthermore, this bespoke two-stage random forest pipeline was never run on plaintext Fashion-MNIST, so it cannot measure information loss; the same section reports the unmodified off-the-shelf MLP dropping from 0.8854 to 0.8188 on HbHAI-protected data.

full rationale

The paper is mostly an empirical black-box evaluation of a proprietary hash-based homomorphic scheme, and several results are legitimate before/after measurements: Dataset 1 clustering and k-NN outputs are compared directly between plaintext and HbHAI-protected forms, data size reduction and speed-up are measured, and the security analysis is explicitly limited to a black-box attempt without claiming a proof. The main circularity is confined to the Fashion-MNIST evidence for the central 'no information loss / unmodified off-the-shelf algorithms' claim. The paper first shows an unmodified MLP degrading from 0.8854 to 0.8188 accuracy, then replaces it with a custom two-stage random forest using an 'empiric decision threshold' and reports 0.9534 as the validation accuracy. That figure is then ranked against 38 plaintext published results and used in the conclusion to confirm that models are 'very well preserved.' Because the threshold was chosen on the validation set whose accuracy is then reported, and because the custom pipeline was never evaluated on plaintext, the headline preservation result reduces to an in-sample fit rather than an independent prediction. The author's dual role as inventor and evaluator, plus citation to his own formalization [12], are disclosure and independence concerns but are not by themselves load-bearing circularity; the Dataset 1 empirical comparisons provide some independent content. Score 6 reflects one central fitted-input-called-prediction step with partial circularity in the paper's central claim.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The central claim rests entirely on the unproven existence of a similarity-preserving, cryptographically secure keyed hash family. The evaluation introduces fitted parameters (empiric decision threshold, M2 model composition) to obtain the reported high accuracy. The security conclusion assumes that absence of observed correlation implies security, which the paper itself concedes is not rigorous. The invented entity 'HbHAI' has no independent evidence.

free parameters (2)
  • class-6 decision threshold (delta=3 Fashion-MNIST) = not disclosed (described as 'empiric decision threshold')
    In Section 4.2 (Analysis delta=3), a second random forest model decides class 6 using a threshold chosen empirically on the HbHAI data. This threshold is load-bearing for the reported 95.34% accuracy, but it is never specified.
  • M2 error-checking model for classes 2, 4, 6 (delta=6) = trained model, parameters undisclosed
    In the delta=6 analysis, a second model M2 is set up specifically to decide classes 2, 4 and 6 after a first model M1. The M1+M2 composition is tuned post hoc and produces the 93.8% figure, but its parameters and decision rules are not disclosed.
assumptions (3)
  • domain assumption There exists a keyed hash function H_{K,delta} satisfying the similarity-order preservation of Definition 1 and the security properties of Definition 2.
    This is the entire foundation of HbHAI. The paper cites [12] (the author's own unpublished specification) and provides no proof. The evaluation cannot test this because the implementation is a black box.
  • domain assumption The two proprietary datasets are representative of 'most AI problems' and permit generalization to the broad claim that no information loss occurs.
    The paper generalizes from one cybersecurity dataset and one image dataset to a general conclusion, with no statistical basis for representativeness.
  • domain assumption Black-box absence of detected matching between plaintext and protected records is evidence of security.
    Section 4.2 and Section 5 treat failure to associate records as a security result, even though the paper itself acknowledges that only a white-box analysis can rigorously assess security. Absence of evidence is presented as evidence of security.
invented entities (1)
  • HbHAI keyed hash function H_{K,delta}
    purpose: Protect data while preserving similarity structure for unmodified AI algorithms.
    The implementation and specification are proprietary and not publicly verifiable. No falsifiable prediction is made outside the vendor's own datasets, and no public API, formal proof, or public dataset is provided to allow outsiders to check it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Technical Evaluation of a Disruptive Approach in Homomorphic AI." pith.science (2026). https://pith.science/paper/ZR6ECMPV

@misc{pith2026250611954,
  author       = {Pith},
  title        = {Pith review of: Technical Evaluation of a Disruptive Approach in Homomorphic AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZR6ECMPV}},
  note         = {Machine review of arXiv:2506.11954}
}
read the original abstract

We present a technical evaluation of a new, disruptive cryptographic approach to data security, known as HbHAI (Hash-based Homomorphic Artificial Intelligence). HbHAI is based on a novel class of key-dependent hash functions that naturally preserve most similarity properties, most AI algorithms rely on. As a main claim, HbHAI makes now possible to analyze and process data in its cryptographically secure form while using existing native AI algorithms without modification, with unprecedented performances compared to existing homomorphic encryption schemes. We tested various HbHAI-protected datasets (non public preview) using traditional unsupervised and supervised learning techniques (clustering, classification, deep neural networks) with classical unmodified AI algorithms. This paper presents technical results from an independent analysis conducted with those different, off-the-shelf AI algorithms. The aim was to assess the security, operability and performance claims regarding HbHAI techniques. As a results, our results confirm most these claims, with only a few minor reservations.

Figures

Figures reproduced from arXiv: 2506.11954 by the authors.

Figure 1
Figure 1. Comparison of learning curves with [16, Chap. 10] algorithms. Plaintext dataset (left), HbHAI-protected version (right) Whatever the method we used, we have observed that class 6 of the dataset was generally badly predicted. So we set up a 2-step detection algorithm based on decision trees [16, Chap. 6] and Random Forests [16, Chap. 7] from scikit.learn. We have proceeded as follows: – Step 1.- We used a Random Fore… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 35 canonical work pages

  1. [12]

    New Proposal for Homomorphic AI

    Eric Filiol (2025). New Proposal for Homomorphic AI. International Conference on the AI Revolution: Research, Ethics, and Society (AIR-RES 2025), April 14th-16th, 2025, Las Vegas, proceedings to be published by Springer

    work page 2025

  2. [1]

    Springer Nature Switzerland AG

    Allon Adir, Ehud Aharoni, Nir Drucker, Ronen Levy, Hayim Shaul and Omri Soceanu (2024).Homomorphic Encryption for Data Science (HE4DS). Springer Nature Switzerland AG

    2024

  3. [2]

    Leveraging TenSEAL: A Comparative Study of BFV and CKKS Schemes for Training ML Models on Encrypted IoT Data

    Fono Louis Aimé, Yancho Basil Wiryen, Noumsi Auguste Vigny and Mvogo Joseph Ngono (2024). Leveraging TenSEAL: A Comparative Study of BFV and CKKS Schemes for Training ML Models on Encrypted IoT Data. Int. J. Inf. Sec. Priv., 18(1), pp. 1–17, https://doi.org/10.4018/IJISP.356402

    work page doi:10.4018/ijisp.356402 2024

  4. [3]

    Tenseal: A library for encrypted tensor operations using homomorphic encryption

    Ayoub Benaissa, Bilal Retiat, Bogdan Cebere and Alaa Eddine Belfedhal (2021). Tenseal: A library for encrypted tensor operations using homomorphic encryption. ArXiv preprint arXiv:2104.03152. https://arxiv.org/abs/2104.03152 Last accessed January 24th

    arXiv 2021

  5. [4]

    Homomor- phic Encryption for Arithmetic of Approximate Numbers

    Jung Hee Cheon, Andrey Kim, Miran Kim, and Yong Soo Song (2017). Homomor- phic Encryption for Arithmetic of Approximate Numbers. In: 23rdrd International Conference on the Theory and Applications of Cryptology and Information Secu- rity (AsiaCrypt), Tsuyoshi Takagi and Thomas Peyrin (Eds.). Lecture Notes in Computer Science 10625, Springer Verlag, pp. 409–437

    work page 2017

  6. [5]

    TFHE: Fast Fully Homomorphic Encryption Library

    Ilaria Chillotti, Nicolas Gama, Mariya Georgieva and Malika Izabachène (2016). TFHE: Fast Fully Homomorphic Encryption Library. Available on https://tfhe.g ithub.io/tfhe/

    work page 2016

  7. [6]

    Attacking FHE-based applications by software fault injections

    Ilaria Chillotti, Nicolas Gama and Louis Goubin (2016). Attacking FHE-based applications by software fault injections. Cryptology ePrint Archive, Paper 2016/1164, https://eprint.iacr.org/2016/1164. Last accessed April 16th, 2025

    work page 2016

  8. [7]

    TFHE: Fast Fully Homomorphic Encryption over the Torus

    Ilaria Chillotti, Nicolas Gama, Mariya Georgieva and Malika Izabachène (2020). TFHE: Fast Fully Homomorphic Encryption over the Torus. Journal of Cryptology, volume 33, pages 34–91. Also available on https://eprint.iacr.org/2018/421

    work page 2020

Show all 42 references
  1. [8]

    A Survey on Implementations of Homomorphic Encryption Schemes

    Thi Van Thao Doan, Mohamed-Lamine Messai, Gérald Gavin and Jérôme Dar- mont (2022). A Survey on Implementations of Homomorphic Encryption Schemes. PREPRINT (Version 2) available at Research Square https://doi.org/10.21203/r s.3.rs-2018739/v2. Last accessed January 24th, 2025. ...

    2022 doi

  2. [9]

    Nonmalleable Cryptography

    Danny Dolev, Cynthia Dwork, Moni Naor (2000). Nonmalleable Cryptography. SIAM Journal on Computing.30 (2), 391-–437

    2000

  3. [10]

    Performance compar- ison of homomorphic encryption scheme implementations

    Goran Ðorđević, Milan Marković and Pavle Vuletić (2021). Performance compar- ison of homomorphic encryption scheme implementations. In: Proceedings of the IcETRAN 2021, pp. 514–520, available on https://www.etran.rs/2021/zbornik/ Papers/104_RTI_2.5.pdf, last accessed December 12th

    2021

  4. [11]

    Somewhat Practical Fully Homo- morphic Encryption

    Junfeng Fan and Frederik Vercauteren (2012). Somewhat Practical Fully Homo- morphic Encryption. Cryptology ePrint Archive, Paper 2012/144. https://eprint .iacr.org/2012/144https://eprint.iacr.org/2012/144. Last accessed January 20th, 2025

    2012

  5. [13]

    Technical Evaluation Hash-based Homomorphic AI

    Eric Filiol (2025). Technical Evaluation Hash-based Homomorphic AI. To be pre- sented at CyberWiseCon 2025 https://cyberwisecon.eu/, May 20 th-23rd, 2025, Vilnius, Lithuania

    2025

  6. [14]

    Evelyn Fix and Joseph L. Hodges. Discriminatory Analysis, Non-parametric Dis- crimination: Consistency Properties. Technical Report 4, USAF School of Aviation Medicine, Randolph Field, 1951

    1951

  7. [15]

    A fully homomorphic encryption scheme

    Craig Gentry (2009). A fully homomorphic encryption scheme. In: Proceedings of the 41st Annual ACM Symposium on Theory of Computing. ACM, pp. 169–178

    2009

  8. [16]

    Hands-On Machine Learning with Scikit-Learn, Keras & Tensor- Flow

    Aurélien Géron. Hands-On Machine Learning with Scikit-Learn, Keras & Tensor- Flow. Third ed., O’Reilly, 2022

    2022

  9. [17]

    The GNU Multiple Precision Arithmetic Library

    Torbjörn Granlund. The GNU Multiple Precision Arithmetic Library. Edition 6.3.0, gmplib.org, 2023

    2023

  10. [18]

    https://github.com/homenc/HElib

    Homenc (2019).HElib: An open-source library implementing homomorphic encryp- tion. https://github.com/homenc/HElib. Last accessed January 24th

    2019

  11. [19]

    https://github.com/hpdic/openfhe-development

    HPDIC Lab (2023).OpenFHE - Open-Source Fully Homomorphic Encryption Li- brary v1.0.3. https://github.com/hpdic/openfhe-development. Last accessed January 24th

    2023

  12. [20]

    Extensions to the k-Means Algorithm for Clustering Large Data Sets with Categorical Values

    Zhexue Huang. Extensions to the k-Means Algorithm for Clustering Large Data Sets with Categorical Values. Data Mining and Knowledge Discovery 2, 283–304 (1998). https://doi.org/10.1023/A:1009769707641

    1998 doi

  13. [21]

    https://www

    International Energy Agency: World Energy Outlook 2024 (2024). https://www. iea.org/reports/world-energy-outlook-2024, last accessed 2025/01/02

    2024

  14. [22]

    https://digi talprivacy.ieee.org/publications/topics/what-is-homomorphic-encryption, last accessed 2025/01/02

    IEEE Digital Privacy: What Is Homomorphic Encryption? (2024). https://digi talprivacy.ieee.org/publications/topics/what-is-homomorphic-encryption, last accessed 2025/01/02

    2024

  15. [23]

    Faster homomorphic comparison opera- tionsforBGVandBFV.author=,ProceedingsonPrivacyEnhancingTechnologies, volume 2021, pp

    Ilia Iliashenko and Vincent Zucca (2021). Faster homomorphic comparison opera- tionsforBGVandBFV.author=,ProceedingsonPrivacyEnhancingTechnologies, volume 2021, pp. 246–264

    2021

  16. [24]

    Introduction to Algebra

    Aleksei Ivanovich Kostrikin (1982). Introduction to Algebra. Universitext Serie, Springer Verlag, New York, Heidelberg, Berlin

    1982

  17. [25]

    FHEBench: Benchmarking Fully Homomorphic En- cryption Schemes

    Lei Jiang and Lei Ju (2022). FHEBench: Benchmarking Fully Homomorphic En- cryption Schemes. ArXiv preprint 2203.00728, https://arxiv.org/abs/2203.00728, last accessed April 15th, 2025

    2022 arXiv

  18. [26]

    A High-Level Technical Overview of Fully Homomorphic En- cryption

    Jeremy Kun (2024). A High-Level Technical Overview of Fully Homomorphic En- cryption. https://www.jeremykun.com/2024/05/04/fhe-overview/, last accessed December 14th, 2024. Technical Evaluation of a Disruptive Approach in Homomorphic AI 15

    2024

  19. [27]

    The MNIST Database of Handwritten Digits

    Yann LeCun, Corinna Cortes and Chris Burges (1998). The MNIST Database of Handwritten Digits. https://yann.lecun.com/exdb/mnist/, last accessed 2024/11/25

    1998

  20. [28]

    Basic Category Theory

    Tom Leinster (2016). Basic Category Theory. Cambridge Studies in Advanced Mathematics, Cambridge University Press. A preprint version is available on https://arxiv.org/abs/1612.09375

    2016 arXiv

  21. [29]

    Wilfred W. K. Lin (2023). Challenges of Homomorphic Encryption. https://www. researchgate.net/publication/370050235_Challenges_of_Homomorphic_encrypt ion, last accessed 2024/11/23

    2023

  22. [30]

    Matheus P. Lobo. Homomorphisms: A Concise Approach. OSF Preprints, osf.io/7 pgcs_v1, 2020

    2020

  23. [31]

    On ideal lattices and learning with errors over rings

    Vadim Lyubashevsky, Chris Peikert, and Oded Regev (2010). On ideal lattices and learning with errors over rings. In: Advances in Cryptology – EUROCRYPT 2010. Lecture Notes in Computer Science 6110, Springer, Berlin, Heidelberg, pp. 1–23

    2010

  24. [32]

    What Is a Statistical Model? The Annals of Statistics, Vol

    Peter McCullagh (2002). What Is a Statistical Model? The Annals of Statistics, Vol. 30, No. 5, pp. 1225-1267

    2002

  25. [33]

    Menezes, Paul C

    Alfred J. Menezes, Paul C. van Oorschot and Scott A. Vanstone (2007).Handbook of Applied Cryptography. CRC Press. Available at http://www.cacr.math.uwater loo.ca/hac/

    2007

  26. [34]

    https: //github.com/microsoft/SEAL

    Microsoft (2018).Microsoft SEAL (Simple Encrypted Arithmetic Library). https: //github.com/microsoft/SEAL. Last accessed January 24th, 2025

    2018

  27. [35]

    US Data Breach Victim Numbers Surge 1170 % Annually

    Paul Muncaster (2024). US Data Breach Victim Numbers Surge 1170 % Annually. Info Security Magazine (2024). https://www.infosecurity-magazine.com/news/u s-data-breach-victims-surge-1170/, last accessed 2024/12/19

    2024

  28. [36]

    William M. Rand. Objective Criteria for the Evaluation of Clustering Methods. Journal of the American Statistical Association Vol. 66, No. 336, pp. 846-850, 1971

    1971

  29. [37]

    Awesome Homomorphic Encryption

    Jonas Schneider (2024). Awesome Homomorphic Encryption. https://github.com /jonaschn/awesome-he, last accessed 2025/01/02

    2024

  30. [38]

    Jaagup Sepp. Datasets for Hash-based Homomorphic AI with Variable Compres- sionRate.WorldCongressinComputerScience,ComputerEngineering,&Applied Computing (CSCE’25) - The 27th International Conference on Artificial Intelli- gence (ICAI’25), July 21st-24th, 2025, Las Vegas, USA....

    2025

  31. [39]

    Comparison of FHE Schemes and Libraries for Efficient Cryptographic Processing

    Arisa Tsuji and Masato Oguchi (2024). Comparison of FHE Schemes and Libraries for Efficient Cryptographic Processing. In: Proceedings of the 2024 International Conference on Computing, Networking and Communications (ICNC): Edge Com- puting, Cloud Computing and Big Data, pp. 58...

    2024

  32. [40]

    On Side-Channel and CVO Attacks against TFHE and FHEW

    Michael Walter (2022). On Side-Channel and CVO Attacks against TFHE and FHEW. Cryptology ePrint Archive, Paper 2022/1722, https://eprint.iacr.org/20 22/1722. Last accessed April 15th, 2025

    2022

  33. [41]

    https://github.com/zalandoresearch/f ashion-mnist, last accessed 2024/11/05

    Zalando Research: Fashion-MNIST (2017). https://github.com/zalandoresearch/f ashion-mnist, last accessed 2024/11/05

    2017

  34. [42]

    Hades: Homomorphic Augmented Decryption for Efficient Symbol-comparison – A Database’s Perspective

    Dongfang Zhao (2024). Hades: Homomorphic Augmented Decryption for Efficient Symbol-comparison – A Database’s Perspective. ArXiv preprint 2412.19980, https: //arxiv.org/abs/2412.19980, last accessed January 24th, 2025

    2024 arXiv

Pith tools

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