pith. sign in

arxiv: 2606.05078 · v1 · pith:MSLAIK2Rnew · submitted 2026-06-03 · 💻 cs.CR

Attention-Augmented LSTMs for Automatic Homophonic Ciphertext Decipherment

Micaella Bruton , Meriem Beloucif , Be\'ata Megyesi This is my paper

Pith reviewed 2026-06-28 05:25 UTC · model grok-4.3

classification 💻 cs.CR
keywords homophonic substitution ciphersattention-augmented LSTMautomatic deciphermentshared-key settinghistorical cryptographysynthetic ciphertextstranscription errors
0
0 comments X

The pith

An attention-augmented LSTM achieves near-perfect character-level decryption of homophonic ciphers when trained only on aligned pairs from a shared code pool.

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

The paper evaluates an attention-augmented LSTM for learning to decrypt homophonic substitution ciphers in a shared-key setting where all texts use subsets of the same code pool. Synthetic ciphertexts are created from historical English and Swedish texts spanning 1500 to 1899, and the model trains solely on aligned pairs without external aids. Results indicate near-perfect accuracy on character level for various lengths, periods, and noisy conditions, with predictable failure outside the pool.

Core claim

The attention-augmented LSTM, trained exclusively on aligned ciphertext-plaintext pairs, learns consistent mappings from multiple codes to each plaintext letter within the shared pool and reaches near-perfect character-level decryption accuracy across English and Swedish texts from 1500-1899, including short examples and those with simulated transcription errors, while failing predictably on texts outside the shared pool.

What carries the argument

Attention-augmented LSTM trained end-to-end on aligned ciphertext-plaintext pairs to learn code-to-letter mappings within a shared homophonic pool.

If this is right

  • Near-perfect accuracy holds across both English and Swedish and across all centuries from 1500 to 1899.
  • Accuracy remains high for short ciphertexts and those containing simulated transcription errors.
  • The model functions as a practical tool for decipherment and for verifying suspected key reuse by failing on texts outside the shared pool.
  • No external language models, frequency statistics, or key-search heuristics are required.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The failure mode could help historians group real historical ciphertexts that likely share a common code pool.
  • The method might reduce dependence on manual analysis when multiple related documents are suspected to use subsets of one key.
  • Applying the approach to actual undeciphered texts would first require constructing or hypothesizing a plausible shared code pool from known historical examples.

Load-bearing premise

All ciphertexts draw from the same known homophonic code pool while each key uses a different but consistent subset of that pool.

What would settle it

Training and testing the model on ciphertexts generated from two entirely separate homophonic code pools and checking whether accuracy stays near-perfect or drops sharply would settle whether the shared-pool condition is what enables the reported performance.

Figures

Figures reproduced from arXiv: 2606.05078 by Be\'ata Megyesi, Meriem Beloucif, Micaella Bruton.

Figure 1
Figure 1. Figure 1: Aligned excerpt from a homophonic digit cipher example from the English test dataset. Each [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the homophonic decryption model. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Homophonic substitution ciphers replace each plaintext letter with one of several possible ciphertext codes, deliberately weakening letter-frequency patterns and making automated decipherment difficult. This paper evaluates whether an attention-augmented Long Short-Term Memory (LSTM) model can learn such mappings in a historically motivated shared-key setting: all ciphertexts draw from the same known homophonic code pool, while individual keys use different consistent subsets of that pool. Using synthetic ciphertexts generated with ChronoFidelius from historical English and Swedish texts dated 1500--1899, we test performance across ciphertext lengths, centuries, variable-length codes, and simulated transcription errors. Models are trained only on aligned ciphertext--plaintext pairs, without external language models, frequency statistics, or key-search heuristics. Results show near-perfect character-level decryption accuracy across both languages and all periods, including short and noisy ciphertexts. The model also fails predictably on ciphertexts outside the shared pool, indicating that it functions as a practical tool for decipherment and key-space verification when key reuse is suspected.

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

1 major / 0 minor

Summary. The manuscript proposes an attention-augmented LSTM model for deciphering homophonic substitution ciphers under a shared-key assumption, where all ciphertexts draw from the same known homophonic code pool but use different consistent subsets for individual keys. Synthetic ciphertexts are generated via ChronoFidelius from historical English and Swedish texts (1500--1899); the model is trained exclusively on aligned ciphertext--plaintext pairs without external language models, frequency data, or key-search heuristics. The central claim is near-perfect character-level decryption accuracy across languages, periods, ciphertext lengths, variable-length codes, and simulated transcription errors, together with predictable failure on ciphertexts drawn from outside the shared pool.

Significance. If the results hold, the work supplies a concrete supervised baseline for learning homophonic mappings directly from aligned pairs and demonstrates that the model can serve as a practical tool for both decipherment and key-space verification when key reuse is suspected. The controlled synthetic regime, out-of-pool negative controls, and coverage of short/noisy texts constitute a falsifiable test of the shared-pool hypothesis.

major comments (1)
  1. Abstract: the claim of 'near-perfect character-level decryption accuracy' is presented without any quantitative metrics, error bars, dataset sizes, number of training examples, or ablation results, rendering the central empirical claim unverifiable from the provided summary and undermining assessment of its load-bearing strength.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the abstract requires quantitative support for its central claims and will revise it in the next version.

read point-by-point responses

  1. Referee: Abstract: the claim of 'near-perfect character-level decryption accuracy' is presented without any quantitative metrics, error bars, dataset sizes, number of training examples, or ablation results, rendering the central empirical claim unverifiable from the provided summary and undermining assessment of its load-bearing strength.

    Authors: We accept this criticism. While the body of the manuscript reports specific metrics (character-level accuracies exceeding 99% on held-out test sets, training on 50,000+ aligned pairs per language/period, dataset sizes by century and length, and ablation studies on attention vs. baseline LSTM), the abstract uses only the qualitative phrase 'near-perfect.' We will revise the abstract to include representative quantitative results, including mean accuracy with standard deviation, number of training examples, and key dataset statistics. This change will be made in the resubmission. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical ML evaluation is self-contained

full rationale

The paper reports supervised training of an attention-augmented LSTM on aligned ciphertext-plaintext pairs generated synthetically from historical texts, followed by accuracy measurements on held-out lengths, noise levels, and out-of-pool controls. No equations, uniqueness theorems, or parameter-fitting steps are described that would make any reported accuracy a definitional consequence of the training regime itself. The shared-pool assumption is an explicit experimental condition whose consequences are tested by the out-of-pool failure case, rather than an unexamined premise smuggled into the result. The work contains no self-citation load-bearing claims or ansatz smuggling; the central result is an empirical observation on a controlled synthetic task.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Central claim rests on the domain assumption of a shared known code pool and aligned training pairs in a synthetic setting; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption All ciphertexts draw from the same known homophonic code pool with individual keys using different consistent subsets.
    Explicitly stated in abstract as the historically motivated shared-key setting required for the model to learn mappings.

pith-pipeline@v0.9.1-grok · 5719 in / 1217 out tokens · 31711 ms · 2026-06-28T05:25:15.965784+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

68 extracted references · 51 canonical work pages · 1 internal anchor

  1. [1]

    doi:10.18653/v1/2025.acl-long.1377 , abstract =

    Adilazuarda, Muhammad Farid and Wijanarko, Musa Izzanardi and Susanto, Lucky and Nur'aini, Khumaisa and Wijaya, Derry Tanti and Aji, Alham Fikri , editor =. doi:10.18653/v1/2025.acl-long.1377 , abstract =

    work page doi:10.18653/v1/2025.acl-long.1377 2025

  2. [2]

    Ahmadzadeh, Ezat and Kim, Hyunil and Jeong, Ongee and Kim, Namki and Moon, Inkyu , year = 2022, journal =. A. doi:10.1109/ACCESS.2022.3140342 , abstract =

    work page doi:10.1109/access.2022.3140342 2022

  3. [3]

    Ahmadzadeh, Ezat and Kim, Hyunil and Jeong, Ongee and Moon, Inkyu , year = 2021, journal =. A. doi:10.1109/ACCESS.2021.3074268 , abstract =

    work page doi:10.1109/access.2021.3074268 2021

  4. [4]

    doi:10.1007/978-981-97-0641-9_51 , abstract =

    Cryptology and. doi:10.1007/978-981-97-0641-9_51 , abstract =

    work page doi:10.1007/978-981-97-0641-9_51

  5. [5]

    Automated

    Antal, Eugen and Mar. Automated. Tatra Mountains Mathematical Publications , volume =. doi:10.2478/tmmp-2022-0019 , urldate =

    work page doi:10.2478/tmmp-2022-0019 2022

  6. [6]

    Encrypted

    Antal, Eugen and Mar. Encrypted. International. doi:10.3384/ecp195689 , abstract =

    work page doi:10.3384/ecp195689

  7. [7]

    Bagla, Kartikay and Gupta, Shivam and Kumar, Ankit and Gupta, Anuj , year = 2024, pages =. Noisy. International. doi:10.1145/3639856.3639889 , abstract =

    work page doi:10.1145/3639856.3639889 2024

  8. [8]

    , year = 2023, journal =

    Bauer, Craig P. , year = 2023, journal =. The. doi:10.1080/01611194.2023.2170158 , langid =

    work page doi:10.1080/01611194.2023.2170158 2023

  9. [9]

    Bonmann, Svenja and Halfmann, Jakob and Korobzow, Natalie and Bobomulloev, Bobomullo , year = 2023, journal =. A. doi:10.1111/1467-968X.12269 , abstract =

    work page doi:10.1111/1467-968x.12269 2023

  10. [10]

    Willis and Kelley, Kathryn and Sarkar, Anoop , editor =

    Born, Logan and Monroe, M. Willis and Kelley, Kathryn and Sarkar, Anoop , editor =. Learning the. Workshop on. doi:10.18653/v1/2023.cawl-1.11 , abstract =

    work page doi:10.18653/v1/2023.cawl-1.11 2023

  11. [11]

    Cryptographic

    Bose, Amrita and Pal, Debranjan and Roy Chowdhury, Dipanwita , editor =. Cryptographic. Applications and. doi:10.1007/978-981-97-9743-1_4 , abstract =

    work page doi:10.1007/978-981-97-9743-1_4

  12. [12]

    Braovi. A. Computational Linguistics , volume =. doi:10.1162/coli_a_00514 , abstract =

    work page doi:10.1162/coli_a_00514

  13. [13]

    Conference on

    Bruton, Micaella and Beloucif, Meriem , editor =. Conference on. doi:10.18653/v1/2023.emnlp-main.671 , abstract =

    work page doi:10.18653/v1/2023.emnlp-main.671 2023

  14. [14]

    Bruton, Micaella and Megyesi, Beata , year = 2025, publisher =. From. International

    2025

  15. [15]

    Computational

    Corazza, Michele , year = 2023, month = jun, address =. Computational. doi:10.48676/unibo/amsdottorato/10975 , abstract =

    work page doi:10.48676/unibo/amsdottorato/10975 2023

  16. [16]

    D, Renny Harlin and Gupta, Kritesh Kumar , year = 2024, pages =. Neural. doi:10.1109/RAICS61201.2024.10690135 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/raics61201.2024.10690135 2024

  17. [17]

    Desenclos, Camille and Lasry, George , year = 2024, publisher =. An. International. doi:10.58009/aere-perennius0090 , abstract =

    work page doi:10.58009/aere-perennius0090 2024

  18. [18]

    and Stamp, Mark , year = 2013, journal =

    Dhavare, Amrapali and Low, Richard M. and Stamp, Mark , year = 2013, journal =. Efficient. doi:10.1080/01611194.2013.797041 , abstract =

    work page doi:10.1080/01611194.2013.797041 2013

  19. [19]

    Fornmark, Filip , year = 2022, address =. Models,

    2022

  20. [20]

    Evaluating

    F. Evaluating. International. doi:10.3384/ecp188394 , abstract =

    work page doi:10.3384/ecp188394

  21. [21]

    and Naughton, Thomas J

    Gopinathan, Unnikrishnan and Monaghan, David S. and Naughton, Thomas J. and Sheridan, John T. , year = 2006, journal =. A. doi:10.1364/OE.14.003181 , abstract =

    work page doi:10.1364/oe.14.003181 2006

  22. [22]

    Automatic

    Han, Songming and Ling, Ying and Xie, Ming and Ming, Shaofeng , year = 2024, pages =. Automatic. doi:10.1109/DSC63484.2024.00073 , abstract =

    work page doi:10.1109/dsc63484.2024.00073 2024

  23. [23]

    Supporting

    H. Supporting. International. doi:10.58009/aere-perennius0100 , abstract =

    work page doi:10.58009/aere-perennius0100

  24. [24]

    Transcription of

    Johansson, Kajsa , year = 2019, address =. Transcription of

    2019

  25. [25]

    Augmented

    Kambhatla, Nishant , year = 2024, address =. Augmented

    2024

  26. [26]

    Kimura, Hayato and Emura, Keita and Isobe, Takanori and Ito, Ryoma and Ogawa, Kazuto and Ohigashi, Toshihiro , editor =. Output. Applied. doi:10.1007/978-3-031-16815-4_15 , isbn =

    work page doi:10.1007/978-3-031-16815-4_15

  27. [27]

    Deciphering an

    Kir. Deciphering an. Cryptologia , volume =. doi:10.1080/01611194.2024.2334042 , abstract =

    work page doi:10.1080/01611194.2024.2334042 2024

  28. [28]

    Knight, Kevin and Megyesi, Be. The. Proceedings of the 4th

  29. [29]

    Cryptanalysis of

    Kopal, Nils , year = 2019, publisher =. Cryptanalysis of. International

    2019

  30. [30]

    Deciphering

    Kopal, Nils and Waldisp. Deciphering. Cryptologia , volume =. doi:10.1080/01611194.2020.1858370 , abstract =

    work page doi:10.1080/01611194.2020.1858370 2020

  31. [31]

    Kopal, Nils , year = 2018, series =. Solving. International

    2018

  32. [32]

    Deciphering

    Lasry, George and Biermann, Norbert and Tomokiyo, Satoshi , year = 2023, journal =. Deciphering. doi:10.1080/01611194.2022.2160677 , abstract =

    work page doi:10.1080/01611194.2022.2160677 2023

  33. [33]

    Deciphering

    Lasry, George and Megyesi, Be. Deciphering. Cryptologia , volume =. doi:10.1080/01611194.2020.1755915 , abstract =

    work page doi:10.1080/01611194.2020.1755915 2020

  34. [34]

    Deciphering a

    Lasry, George and Paolo, Bonavoglia , year = 2022, publisher =. Deciphering a. International. doi:10.3384/ecp188401 , abstract =

    work page doi:10.3384/ecp188401 2022

  35. [35]

    Applying

    Lehofer, Anna , year = 2022, journal =. Applying. doi:10.1080/01611194.2021.1918801 , abstract =

    work page doi:10.1080/01611194.2021.1918801 2022

  36. [36]

    Decrypting

    Lehofer, Anna , year = 2020, journal =. Decrypting. doi:10.3311/ope.381 , urldate =

    work page doi:10.3311/ope.381 2020

  37. [37]

    Detection of

    Leierzopf, Ernst and Mikhalev, Vasily and Kopal, Nils and Esslinger, Bernhard and Lampesberger, Harald and Hermann, Eckehard , editor =. Detection of. Data. doi:10.1007/978-981-16-8531-6_11 , abstract =

    work page doi:10.1007/978-981-16-8531-6_11

  38. [38]

    Deciphering

    Luo, Jiaming and Hartmann, Frederik and Santus, Enrico and Barzilay, Regina and Cao, Yuan , year = 2021, journal =. Deciphering. doi:10.1162/tacl_a_00354 , abstract =

    work page doi:10.1162/tacl_a_00354 2021

  39. [39]

    Cryptography

    Mardon, Austin and Barara, Gurman and Chana, Ishpreet and DiMartino, Alexia and Falade, Irene and Harum, Rokya and Hauser, Alexandra and Johnson, Jamie and Li, Amy and Pham, Jennifer and Varghese, Noah , year = 2021, publisher =. Cryptography

    2021

  40. [40]

    Benchmarking

    Maskey, Utsav and Zhu, Chencheng and Naseem, Usman , editor =. Benchmarking. Empirical. doi:10.18653/v1/2025.findings-emnlp.1082 , abstract =

    work page doi:10.18653/v1/2025.findings-emnlp.1082 2025

  41. [41]

    Deciphering

    Megyesi, Be. Decryption of. Cryptologia , volume =. doi:10.1080/01611194.2020.1716410 , abstract =

    work page doi:10.1080/01611194.2020.1716410 2020

  42. [42]

    Historical

    Megyesi, B. Historical. International. doi:10.3384/ecp195701 , abstract =

    work page doi:10.3384/ecp195701

  43. [43]

    Transcription of

    Megyesi, Be. Transcription of. International. doi:10.3384/ecp2020171014 , abstract =

    work page doi:10.3384/ecp2020171014

  44. [44]

    Megyesi, Be. What. International. doi:10.3384/ecp188404 , abstract =

    work page doi:10.3384/ecp188404

  45. [45]

    Structured

    M. Structured. European. doi:10.1007/978-3-031-91572-7_20 , abstract =

    work page doi:10.1007/978-3-031-91572-7_20

  46. [46]

    Plaintext

    Najeeb, Mahira and Ghulam, Zikra and Pervaiz, Zahid and Masood, Ammar , year = 2025, journal =. Plaintext. doi:10.1007/s00521-024-10734-w , abstract =

    work page doi:10.1007/s00521-024-10734-w 2025

  47. [47]

    Cryptography:

    Naser, S M , year = 2021, journal =. Cryptography:

    2021

  48. [48]

    Improved

    Nuhn, Malte and Schamper, Julian and Ney, Hermann , year = 2014, pages =. Improved. Empirical. doi:10.3115/v1/D14-1184 , abstract =

    work page doi:10.3115/v1/d14-1184 2014

  49. [49]

    Evolutionary

    Oranchak, David , year = 2008, pages =. Evolutionary. Annual. doi:10.1145/1389095.1389425 , abstract =

    work page doi:10.1145/1389095.1389425 2008

  50. [50]

    Automated

    Park, Seonghwan and Kim, Hyunil and Moon, Inkyu , year = 2023, journal =. Automated. doi:10.3390/cryptography7030035 , urldate =

    work page doi:10.3390/cryptography7030035 2023

  51. [51]

    Pettersson, Eva and Megyesi, Be. The. Digital. doi:10.5617/dhnbpub.11045 , abstract =

    work page doi:10.5617/dhnbpub.11045

  52. [52]

    Matching

    Pettersson, Eva and Megyesi, Beata , editor =. Matching. Nordic

  53. [53]

    Deciphering

    Pourdamghani, Nima and Knight, Kevin , editor =. Deciphering. Empirical. doi:10.18653/v1/D17-1266 , abstract =

    work page doi:10.18653/v1/d17-1266

  54. [54]

    Bayesian

    Ravi, Sujith and Knight, Kevin , editor =. Bayesian. Association for

  55. [55]

    Tumblr , urldate =

    The Fanfare Is in the Light but the Execution Is in the Dark , author =. Tumblr , urldate =

  56. [56]

    Sikora, Justyna , year = 2022, address =. The

    2022

  57. [57]

    International

    Simhadri, Sevitha and. International. doi:10.1109/NETCRYPT65877.2025.11102567 , abstract =

    work page doi:10.1109/netcrypt65877.2025.11102567 2025

  58. [58]

    Srivastava, Shivin and Bhatia, Ashutosh , year = 2018, pages =. On the. doi:10.1109/ICBK.2018.00029 , abstract =

    work page doi:10.1109/icbk.2018.00029 2018

  59. [59]

    Improving

    Stenlund, Mathias Hans Erik and Nanni, Mathilde and Bruton, Micaella and Beloucif, Meriem , editor =. Improving. International

  60. [60]

    Szigeti, Ferenc and H. The. International. doi:10.3384/ecp188409 , abstract =

    work page doi:10.3384/ecp188409

  61. [61]

    Accessing

    Tlachenska, Elina and Ivanov, Kiril and Nenova, Maria and Doynov, Rumen and. Accessing. International. doi:10.1109/ICEST66328.2025.11098206 , abstract =

    work page doi:10.1109/icest66328.2025.11098206 2025

  62. [62]

    BAD SLAM: Bundle adjusted direct RGB-D SLAM,

    Yin, Xusen and Aldarrab, Nada and Megyesi, Be. Decipherment of. International. doi:10.1109/ICDAR.2019.00022 , urldate =

    work page doi:10.1109/icdar.2019.00022 2019

  63. [63]

    Zhong, Qing and Liu, Bo and Ren, Jianxin and Li, Yongxin and Guo, Zhiruo and Mao, Yaya and Wu, Xiangyu and Ma, Yiming and Wu, Yongfeng and Zhao, Lilong and Sun, Tingting and Ullah, Rahat , year = 2022, journal =. Self-. doi:10.1364/OE.468945 , abstract =

    work page doi:10.1364/oe.468945 2022

  64. [64]

    Cryptanalysis on

    Zhou, Rong and Yu, Simin , year = 2025, journal =. Cryptanalysis on. doi:10.1007/s11071-025-11582-6 , langid =

    work page doi:10.1007/s11071-025-11582-6 2025

  65. [65]

    Enhancing

    Bastian, Maik and Esslinger, Bernhard and Hermann, Eckehard and Kopal, Nils and Lampesberger, Harald , year = 2025, eprint =. Enhancing. International

    2025

  66. [66]

    Nada Aldarrab , title =

  67. [67]

    2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=

    Decipherment of historical manuscript images , author=. 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=. 2019 , organization=

    2019

  68. [68]

    Hochreiter, Sepp and Schmidhuber, J. Long. Neural Computation , volume =. doi:10.1162/neco.1997.9.8.1735 , abstract =

    work page doi:10.1162/neco.1997.9.8.1735 1997