See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/365187162

EveGAN: Using Generative Deep Learning for Cryptanalysis

Conference Paper · November 2022

DOI: 10.1145/3548606.3563493

CITATIONS READS
0 140

1 author:

Roger Hallman
CAT Labs (https://catlabs.io/)
52 PUBLICATIONS 217 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

DAES 2017 : The Third International Workshop on Data Analytics and Emerging Services (DAES) with the International Conference on Engineering and Technology 2017 View
project

IoDDoS -- The Internet of Distributed Denial of Service Attacks View project

All content following this page was uploaded by Roger Hallman on 08 November 2022.

The user has requested enhancement of the downloaded file.

 Poster: EveGAN: Using Generative Deep Learning for
Cryptanalysis
Roger A. Hallman∗
Thayer School of Engineering, Dartmouth College
Hanover, New Hampshire, USA &
Naval Information Warfare Center (NIWC) Pacific
San Diego, California, USA
[email protected]

ABSTRACT 1 INTRODUCTION
Cryptography and Machine Learning are two computational sci- Cryptography and machine learning are two computational ca-
ence fields that intuitively seem related. Privacy-preserving ma- pabilities that are foundational to modern life; we interact almost
chine learning–either utilizing encrypted models or learning over unknowingly with each during the course of an average day. Indeed,
encrypted data–is an exploding field thanks to the maturation of while the two fields may seem wholly separate at first blush there is
primitives such as fully homomorphic encryption and secure mul- a long relationship between them [19]. The “Cambrian Explosion"
tiparty computation. However there has been surprisingly little of data due to the ever-increasing adoption of distributed sensor
work on applying recent advances in machine learning to the task networks, the Internet of Things (IoT), mobile and wearable devices,
of cryptanalysis, the branch of cryptography that studies how cryp- and other networked ‘smart’ technologies has led to a public aware-
tographic ciphers can be attacked. In particular, while a crypto- ness of just how much sensitive information they produce. There
graphic cipher seeks to keep certain information secret by making is a natural tension as governments and businesses would like to
it appear random, discerning patterns and structure from random utilize this data while the public would like sensitive data kept
data is a common machine learning task. This poster paper pro- private. Privacy-preserving machine learning (PPML) [13], employ-
poses EveGAN, an approach that treats cryptanalysis as a language ing advanced cryptography such as fully homomorphic encryption
translation problem. While treating cipher cracking as a language (FHE), secure multiparty computation, and differential privacy offer
translation problem has been validated against a handful of clas- the ability to utilize this deluge of data while protecting the sen-
sical substitution ciphers, the EveGAN approach builds on these sitive information of individual data subjects. However, there are
results to create a new class of generative deep learning-based surprisingly few applications of machine learning to cryptanalytic
cryptanalysis attacks. attacks against cryptographic systems.
To that end, this paper proposes EveGAN as an approach to
CCS CONCEPTS cracking cryptographic ciphers that utilizes recent advancements
• Security and privacy → Cryptanalysis and other attacks; • in generative deep learning. The key insight is that deciphering
Computing methodologies → Adversarial learning; Neural encrypted texts can be thought of as a language translation problem
networks. [9]. Furthermore, if a deep learning system can learn to accurately
decipher an encrypted message, then it should be able to mimic the
KEYWORDS encryption algorithm well enough to “encrypt" a fraudulent mes-
sage that can be meaningfully decrypted by encryption algorithm’s
Cryptography, Cryptanalysis, Generative Adversarial Networks,
evaluation function. If successful, the EveGAN approach will open
Language Translation, Forgery/Poisoning Attacks
up new classes of cryptoanalytic attacks.
ACM Reference Format: The remainder of this paper is organized as follows: A terse but
Roger A. Hallman. 2022. Poster: EveGAN: Using Generative Deep Learning sufficient introduction to relevant background information is given
for Cryptanalysis. In Proceedings of 2022 ACM SIGSAC Conference on Com- in Section 2. Section 3 presents the EveGAN approach to generative
puter and Communications Security (CCS ’22). ACM, New York, NY, USA, deep learning-based cryptanalysis. Section 4 presents concluding
3 pages. https://doi.org/10.1145/3548606.3563493
remarks, as well as planned future work.
∗ RogerA. Hallman is partially supported by the United States Department of Defense
SMART Scholarship for Service Program, funded by USD/R&E (The Under Secretary
of Defense-Research and Engineering), National Defense Education Program (NDEP) /
2 BACKGROUND AND RELATED WORK
BA-1, Basic Research.
2.1 Generative Deep Learning and Language
Permission to make digital or hard copies of part or all of this work for personal or Translation
classroom use is granted without fee provided that copies are not made or distributed Deep learning refers to machine learning architectures that are
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored. inspired after human neurological systems [2]. Where many statis-
For all other uses, contact the owner/author(s). tical learning approaches perform well with moderately-sized data
CCS ’22, November 07–11, 2022, Los Angeles, Ca, USA sets, deep learning systems tend towards improved performance
© 2022 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-9450-5/22/11. as the size of the data set increases. Generative deep learning [18]
https://doi.org/10.1145/3548606.3563493 seeks to learn the probability distribution for a class of data and

3355
CCS ’22, November 07–11, 2022, Los Angeles, Ca, USA Roger A. Hallman

Figure 1: The EveGAN Cryptanalysis Framework.

then create synthetic data that is similar to that class. Most of the rely on knowledge of context (e.g., the plaintext language) and uti-
recent advances in generative deep learning have come from the lize tools of frequency analysis to discover the key. While there are
field of generative adversarial networks (GANs) [10], which simulta- many strategies to attack a cipher, this work exclusively considers
neously train generator and discriminator networks to create more the ciphertext-only attack [4] wherein Eve can only capture en-
realistic synthetic data. GANs have led to great advances in image crypted messages. Ciphertext-only attacks have been successfully
processing [24] and captioning [5], as well as language translation leveraged against both classical and modern ciphers.
[22]. Transformers [20], a recently-developed attention-based ar-
chitecture, have led to groundbreaking results in various language 2.3 Deep Learning and Cryptanalysis
processing tasks [11]. GAN architectures utilizing transformers Cryptographic ciphers seek to guarantee data confidentiality by
[23] have shown convincing results with stylistic text generation. making it appear random while machine learning systems are often
tasked with finding patterns and order in seemingly random noise
2.2 Cryptography and Cryptanalysis [1]. There have been surprisingly few applications of advanced ma-
Cryptography is a method of keeping information secret and pro- chine learning techniques to cryptanalysis attacks. Gohr’s seminal
viding a guarantee that only authorized parties should have access paper [7] demonstrated the potential for deep learning, successfully
to it by making that structured data difficult to recognize by appear- recovering keys by implementing a differential cryptanalysis attack
ing to be random. This is distinct from steganography [16], which against the SPECK-32/64 block cipher. More recently, Wenger, et
attempts to hide information by embedding it in some other object al. [21] used a transformer-based architecture to recover secret
(e.g., a digital image, a data stream, etc.). With a cryptographic messages hidden by the Learning with Errors Problem, on which
approach, Alice and Bob would like to share information that they modern FHE schemes are built. While these results are not immedi-
want to keep secret from an adversary, Eve. There are too many ately germane to the approach of this paper, their results illustrate
cryptographic ciphers to mention here and the reader is referred to the possibilities for deep learning and cryptanalysis and may inform
[12] for a detailed exposition on classic and modern cryptography. future work directions. Ahmadzadeh, et al. [3] used a bidirectional
The “gold standards" for modern ciphers are semantic security [8] Recurrent Neural Network model for ciphertext classification from
and indistinguishability [17], equivalent criteria [12] guaranteeing classical substitution ciphers. Most relevant to this paper’s approach
that no adversary has a better than 50% chance of determining are instances where machine learning-based language translation
whether a ciphertext 𝑐 is an encryption of two different plaintext tools were used to decipher ciphertexts from classic substitution
messages 𝑚 1 or 𝑚 2 . Alice and Bob agree on a cryptographic key 𝑘, and homophonic ciphers. Knight’s 2011 paper [14] hints at the po-
then Alice encrypts her message 𝑚 with the function Enc(𝑘, 𝑚) = tential of this approach to produce a reasonable translation of the
𝑐. Bob is able to decrypt 𝑐 with an evaluation function Eval(𝑘, 𝑐) = Copiale Cipher, an enciphered book that dates to the mid-19th Cen-
𝑚. Eve would like to discover 𝑚 and so she must use cryptanalysis. tury. A more recent GAN-based approach [9] utilizes CipherGAN,
Cryptanalysis [6] is the branch of cryptography that studies how an adaptation of the CycleGAN architecture, to decipher encrypted
to attack and break cryptographic ciphers, often with the ultimate texts from Shift and Vigenère Ciphers, two well-known classes of
goal of recovering the cryptographic key. Attacks on classic ciphers substitution cipher.

3356
 Poster: EveGAN: Using Generative Deep Learning for Cryptanalysis CCS ’22, November 07–11, 2022, Los Angeles, Ca, USA

3 THE EVEGAN APPROACH ciphertexts, but would also observe actions taken which correspond
This work proposes the inverse problem presented in [9]: Can a with observed ciphertexts. If it is successful, the EveGAN approach
deep learning system learn to mimic (i.e., learn a representation of) will lead to new classes of cryptanalysis attacks against modern
a cryptographic cipher well enough that it can produce a synthetic cryptographic systems.
ciphertext that can be decrypted by the cipher’s evaluation function?
Moreover, while treating cipher cracking as a language translation REFERENCES
[1] Najwa Aaraj. 2021. How machine learning and cryptography might
problem works for classical substitution ciphers, questions remain work better together. Blockchain Tribune (December 28 2021).
as to whether this approach will work against modern, semantically https://blockchaintribune.com/how-machine-learning-and-cryptography-
secure cryptographic systems. might-work-better-together/.
[2] Charu C Aggarwal et al. 2018. Neural networks and deep learning. Springer 10
The basic EveGAN framework (Figure 1) is laid out as follows: (2018), 978–3.
Alice has a series of messages {𝑚 1, ..., 𝑚𝑖 } that she would like to [3] Ezat Ahmadzadeh, Hyunil Kim, Ongee Jeong, Namki Kim, and Inkyu Moon. 2022.
share only with Bob, therefore she uses a cryptographic cipher to A Deep Bidirectional LSTM-GRU Network Model for Automated Ciphertext
Classification. IEEE Access 10 (2022), 3228–3237.
encrypt her messages and sends Bob the evaluation function so [4] Alex Biryukov and Eyal Kushilevitz. 1998. From differential cryptanalysis to
that he can correctly decrypt Alice’s messages. In keeping with ciphertext-only attacks. In Annual International Cryptology Conference. Springer,
72–88.
Kerkhoff’s Principle [15], an adversary is assumed full knowledge [5] Chen Chen, Shuai Mu, Wanpeng Xiao, Zexiong Ye, Liesi Wu, and Qi Ju. 2019.
of the system except for the encryption key. The adversary observes Improving image captioning with conditional generative adversarial nets. In
the ciphertexts 𝑥𝑟𝑒𝑎𝑙 in transit from Alice to Bob and processes 𝑥𝑟𝑒𝑎𝑙 Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 8142–8150.
[6] Christophe De Canniere, Alex Biryukov, and Bart Preneel. 2006. An introduction
with EveGAN in order to generate a synthetic ciphertext which can to block cipher cryptanalysis. Proc. IEEE 94, 2 (2006), 346–356.
be sent to Bob. [7] Aron Gohr. 2019. Improving attacks on round-reduced speck32/64 using deep
EveGAN takes 𝑥𝑟𝑒𝑎𝑙 and inputs it into two modules, (𝑖) a Dis- learning. In Annual International Cryptology Conference. Springer, 150–179.
[8] Shafi Goldwasser and Silvio Micali. 1984. Probabilistic encryption. Journal of
criminator Network and (𝑖𝑖) a Ciphertext Classifier that extends computer and system sciences 28, 2 (1984), 270–299.
the results from [3]. The ciphertext is taken as representative of [9] Aidan N. Gomez, Sicong Huang, Ivan Zhang, Bryan M. Li, Muhammad Osama,
and Lukasz Kaiser. 2018. Unsupervised Cipher Cracking Using Discrete GANs.
a language. The Discriminator Network is trained to differentiate In International Conference on Learning Representations. https://openreview.net/
between the real and mimicked ciphertext language. The Ciphertext forum?id=BkeqO7x0-
Classifier module is trained on a prototype of Alice’s cipher and [10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial
is used to give the generator network message context 𝑥𝑐𝑙𝑎𝑠𝑠 (e.g., nets. Advances in neural information processing systems 27 (2014).
message topics, numerical ranges, etc.). The Generator Network [11] Shigeki Karita, Nanxin Chen, Tomoki Hayashi, Takaaki Hori, Hirofumi Inaguma,
takes a seed message 𝑧 and outputs a synthetic ciphertext 𝑥 𝑓 𝑎𝑘𝑒 Ziyan Jiang, Masao Someki, Nelson Enrique Yalta Soplin, Ryuichi Yamamoto,
Xiaofei Wang, et al. 2019. A comparative study on transformer vs rnn in speech
which is routed to the Discriminator Network. The Discrimina- applications. In 2019 IEEE Automatic Speech Recognition and Understanding Work-
tor Network should reject 𝑥 𝑓 𝑎𝑘𝑒 until the Generator is sufficiently shop (ASRU). IEEE, 449–456.
[12] Jonathan Katz and Yehuda Lindell. 2020. Introduction to modern cryptography.
trained at mimicking 𝑥𝑟𝑒𝑎𝑙 . The Generator Network will eventually CRC press.
succeed in outputting a synthetic ciphertext 𝑥𝑚𝑖𝑚𝑖𝑐 that fools the [13] Kwangjo Kim and Harry Chandra Tanuwidjaja. 2021. Privacy-preserving Deep
Discriminator Network. EveGAN injects the synthetic ciphertext Learning: A Comprehensive Survey. Springer.
[14] Kevin Knight, Beáta Megyesi, and Christiane Schaefer. 2011. The copiale cipher.
𝑥𝑚𝑖𝑚𝑖𝑐 into the stream that gets routed to Bob for decryption. In Proceedings of the 4th Workshop on Building and Using Comparable Corpora:
Comparable Corpora and the Web. 2–9.
[15] Thorsten Knoll. 2018. Adapting Kerckhoffs’s principle. Advanced Microkernel
3.1 Open Questions Raised by EveGAN Operating Systems (2018), 93.
EveGAN is a framework that utilizes recent advances in generative [16] Wojciech Mazurczyk, Steffen Wendzel, Sebastian Zander, Amir Houmansadr, and
Krzysztof Szczypiorski. 2016. Information hiding in communication networks:
deep learning and language translation for cryptanalysis. This ap- fundamentals, mechanisms, applications, and countermeasures. John Wiley &
proach builds on promising results against classical substitution Sons.
ciphers [9] but a number of open questions remain: (𝑖) How well [17] Moni Naor and Moti Yung. 1990. Public-key cryptosystems provably secure
against chosen ciphertext attacks. In Proceedings of the twenty-second annual
will EveGAN work against modern semantically secure ciphers? (𝑖𝑖) ACM symposium on Theory of computing. 427–437.
How will key switches and subsequent retraining affect EveGAN [18] Achraf Oussidi and Azeddine Elhassouny. 2018. Deep generative models: Survey.
In 2018 International Conference on Intelligent Systems and Computer Vision (ISCV).
performance? (𝑖𝑖𝑖) How effective would EveGAN’s mimicked data IEEE, 1–8.
be for mounting a poisoning attack against PPML learning models? [19] Ronald L Rivest. 1991. Cryptography and machine learning. In International
(𝑖𝑣) Assuming that EveGAN can effectively mimic Alice’s cipher, Conference on the Theory and Application of Cryptology. Springer, 427–439.
[20] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
can this achievement be leveraged into a key recovery attack? Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. Advances in neural information processing systems 30 (2017).
[21] Emily Wenger, Mingjie Chen, François Charton, and Kristin Lauter. 2022.
4 CONCLUSION AND FUTURE WORK SALSA: Attacking Lattice Cryptography with Transformers. arXiv preprint
This paper presents EveGAN, a conceptual cryptanalysis framework arXiv:2207.04785 (2022).
[22] Zhen Yang, Wei Chen, Feng Wang, and Bo Xu. 2018. Generative adversarial
that treats cipher cracking as a language translation problem. As training for neural machine translation. Neurocomputing 321 (2018), 146–155.
it is conceptual there is ongoing work, at present validating the [23] Kuo-Hao Zeng, Mohammad Shoeybi, and Ming-Yu Liu. 2020. Style example-
framework against a classical substitution cipher as a proof-of- guided text generation using generative adversarial transformers. arXiv preprint
arXiv:2003.00674 (2020).
concept. Future work will include testing EveGAN against modern [24] He Zhang, Vishwanath Sindagi, and Vishal M Patel. 2019. Image de-raining using
ciphers. Additionally, it would be interesting to test EveGAN against a conditional generative adversarial network. IEEE transactions on circuits and
systems for video technology 30, 11 (2019), 3943–3956.
an applied cipher such as an encrypted control system. This scenario
would be particularly revealing as EveGAN would not just capture

3357

View publication stats