Empirical Ciphertext Distinguishability Testing of KEMs and Hybrid Encryption via Deep Learning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: I'm Nadia.
Elias: And I'm Elias, here with Priya.
Nadia: We're looking at the paper: "Empirical Ciphertext Distinguishability Testing of KEMs and Hybrid Encryption via Deep Learning".
Elias: Here's the gist of the paper. Ensuring ciphertext indistinguishability is fundamental to cryptographic security, but empirically validating this property in real implementations and hybrid settings presents practical challenges <ref:2604.06942#pg2>.
Nadia: By modeling indistinguishability under chosen-plaintext attack (IND-CPA) games as binary classification tasks and training on labeled ciphertext data with binary cross-entropy loss, we study deep neural network (DNN) distinguishers for ciphertext indistinguishability <ref:2604.06942#pg4>.
Elias: We apply this methodology to PQC key encapsulation mechanisms (KEMs) <ref:2604.06942#pg4>.
Nadia: The main contributions of this work include: 1.
Title and authors: Nadia: To get into the details of this paper, we have to look at how they frame their approach in the title and who is behind it. They're testing KEMs and hybrid encryption using deep learning. It’s a direct connection between modern PQC constructions and machine learning techniques.
Elias: And those authors, Simon Calderon, Niklas Johansson, and Onur Günlü, bring different expertise to the table—one focused on the algorithms themselves, another on proofs and parameters for those algorithms, and one who brings in a perspective on measurement research.
Priya: It’s interesting how they combine these fields because we always need to know if a new construction actually holds up when you look at it through an empirical lens rather than just pure mathematics.
Nadia: They are essentially using this DNN modeling of the IND-CPA game to empirically test the IND-CPA properties of the underlying public key encryption schemes they study, specifically ML-KEM, BIKE, and HQC.
Elias: And what's a real point for us is that they don't just stop at testing individual KEMs; they extend this methodology to test hybridized KEMs as well. This moves the research into a more complex area where combining different primitives matters.
Priya: So it’s about checking security inheritance when you mix things up, which is something we deal with a lot in privacy and measurement studies when looking at composite systems.
Nadia: Right, and their main contribution is presenting this novel extension to test these hybridized KEMs where the construction uses a provably secure combiner of the form k equals F of k1 c XOR F of k2 c, which is quite specific.
Elias: That combination structure, where you take two shared secrets from component KEMs and combine them using XOR with some function F applied to the concatenated ciphertext c, that's a key focus for them. It tests if the resulting shared secret k is still secure under this model.
The paper's summary: Nadia: So what they summarize is that their DNN framework complements formal security analysis by providing a practical tool to evaluate indistinguishability and find potential weaknesses in both implementations and composed constructions. It’s meant to be an empirical supplement to the math.
Elias: They detail their main contributions, starting with using the DNN-based modeling of the IND-CPA game to test individual PQC KEMs like ML-KEM, BIKE, and HQC. That’s testing the security of those standalone components.
Priya: I'm curious about how they handle the data for this training part; it involves training a DNN to distinguish between encryptions of uniformly random plaintexts and encryptions of a fixed plaintext, like all zeros. It’s a standard setup but applied here for security testing.
Nadia: That’s right, they train each DNN using stochastic gradient descent with binary cross-entropy loss on this labeled ciphertext data. They are looking at whether the network can tell the difference between random and fixed inputs based on the ciphertext alone.
Elias: The summary also highlights their extension to test hybrid KEMs, where they input a concatenation of the ciphertext outputs from both components when testing that specific combiner structure we talked about earlier.
Priya: So when you look at hybrid systems, they are feeding the network two parts of the ciphertext and checking if those parts decrypt or decapsulate to the same underlying value or two independent values. It’s a way to check for security inheritance in that specific combination.
The paper's improvements: Nadia: They suggest several ways they improve things, focusing on extending the testing beyond just individual KEMs and into those more complex hybrid scenarios we discussed.
Elias: One major extension is precisely what they present: allowing the testing of hybridized KEMs constructed with that provably secure combiner structure, k equals F of k1 c XOR F of k2 c. It tests security when you mix any KEM with any asymmetric cipher in this way.
Priya: And they also demonstrate the generality by applying their deep learning-based constructions to other hybridized scenarios, such as cascade symmetric encryption, showing that it works beyond just standard KEM combinations.
Nadia: That cascade symmetric encryption testing shows they can apply this methodology to other complex composite constructions too, proving the method has some methodological generality. It’s not just limited to KEM mixing.
Elias: And they also introduce a rigorous statistical analysis for their experiments, moving beyond what you might see in typical deep learning papers. They use two-sided binomial hypothesis testing to determine whether the neural binary classifiers actually gain a nonnegligible advantage over random guessing.
Priya: That statistical rigor is important because it tells us if the observed accuracy improvement is anything meaningful or just noise, and they show that even when an insecure component algorithm like plain RSA is used, the accuracies stay close to fifty percent.
Conclusion: Nadia: So wrapping up the paper Empirical Ciphertext Distinguishability Testing of KEMs and Hybrid Encryption via Deep Learning. They conclude that no algorithm or combination of algorithms demonstrated a significant advantage when tested against this DNN adversary model.
Elias: This means, according to their rigorous statistical analysis using two-sided binomial tests with a significance level alpha equals zero point zero one, there’s no evidence of any exploitable patterns under the considered deep learning adversary model. It aligns with the theoretical guarantees that hybrids including at least one IND-CPA secure component preserve indistinguishability.
Priya: I think what’s important for us is that this method provides an adaptive, practical way to validate indistinguishability in more general settings, which complements traditional analytical security analysis. It gives us a data-driven way to check implementations and compositions.
Nadia: They also point out some limitations of using this DNN framework because the neural classifier doesn't have access to public keys or encryption oracles, which limits the types of attack vectors it can capture directly.
Elias: That’s a fair caveat; while it captures fundamental aspects of the IND-CPA game, if you want to model attacks that require access to those external resources, this specific DNN modeling has its limits. It's an empirical estimator for real-world validation, not a complete security proof in itself.
Priya: So the paper provides a way to use deep learning as an adaptive, practical estimator for indistinguishability in more general IND-CPA settings and helps us validate implementations and compositions empirically. That’s the core value here.
Simon Calderon, Niklas Johansson, Onur Günülue
Linköping University · Technische University Dortmund
cs.CR, cs.IT, cs.LG, cs.NE, eess.SP, math.IT
Submitted: 2026-04-08
Updated: 2026-11-02
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 77/100
The gist: Ensuring ciphertext indistinguishability is fundamental to cryptographic security, but empirically validating this property in real implementations and hybrid settings presents practical challenges…
Key concepts
- IND-CPA Game
- This is a cryptographic game where an attacker tries to distinguish between two types of ciphertexts: one encrypted with a truly random message and another encrypted with a fixed, specific message. The goal is to prove that the ciphertext looks random, making it impossible for an attacker to tell the difference.
- DNN-based Modeling
- The paper uses Deep Neural Networks (DNNs) as a tool to model this IND-CPA game. A DNN is trained on labeled data—ciphertexts known to come from random messages versus fixed messages—to learn patterns that might reveal weaknesses in the underlying encryption scheme.
- Hybrid KEM Combiner
- This refers to combining two separate KEM components, $k = F(k1, c) ext{ ⊕ } F(k2, c)$. The DNN tests if this combination preserves security. It checks if the combined output looks random even when one component is weak or insecure.
- Two-Sided Binomial Hypothesis Testing
- This is a rigorous statistical method used to determine if the performance of a neural classifier (the DNN) is significantly better than random guessing. It helps confirm whether the observed accuracy gains are real or just due to chance.
Terminology
Summary
Ensuring ciphertext indistinguishability is fundamental to cryptographic security, but empirically validating this property in real implementations and hybrid settings presents practical challenges <ref:2604. p. 0 de
Abstract. Ensuring ciphertext indistinguishability is fundamental to
cryptographic security, but empirically validating this property in real
implementations and hybrid settings presents practical challenges. The
transition to post-quantum cryptography (PQC), with its hybrid constructions combining classical and quantum-resistant primitives, makes
empirical validation approaches increasingly valuable. By modeling indistinguishability under chosen-plaintext attack (IND-CPA) games as
binary classification tasks and training on labeled ciphertext data with
binary cross-entropy loss, we study deep neural network (DNN) distinguishers for ciphertext indistinguishability. By modeling indistinguishability under chosen-plaintext attack (IND-CPA) games as
binary classification tasks and training on labeled ciphertext data with
binary cross-entropy loss, we study deep neural network (DNN) distinguishers for ciphertext indistinguishability. We apply this methodology
to PQC key encapsulation mechanisms (KEMs).Where the paper says it
Where the paper says it
Where the paper says it
Main Contributions
The main contributions of this work include:
-
We use a DNN-based modeling of the IND-CPA game to empirically test the IND-CPA properties of the underlying public key encryption (PKE) schemes used in ML-KEM, BIKE, and HQC <ref:2604. p. 2
06942#pg4>.Where the paper says it
The proposed neural-network framework complements formal security analysis by providing a practical and flexible, empirical tool to evaluate indistinguishability and identify potential weaknesses in implementations and composed constructions. The main contributions of this work include: • We use a DNN-based modeling of the IND-CPA game to empirically test the IND-CPA properties of the underlying public key encryption (PKE) schemes used in ML-KEM, BIKE, and HQC.
-
We present a novel extension to this framework to test hybridized KEMs <ref:2604. p. 2
06942#pg4>. Specifically, our extension allows for testing hybridized KEMs, where the KEM is constructed using a provably-secure combiner of the form k = F(k1, c) ⊕ F(k2, c) [16] <ref:2604. p. 2Where the paper says it
The main contributions of this work include: • We use a DNN-based modeling of the IND-CPA game to empirically test the IND-CPA properties of the underlying public key encryption (PKE) schemes used in ML-KEM, BIKE, and HQC. • We present a novel extension to this framework to test hybridized KEMs.
06942#pg4>.Where the paper says it
• We present a novel extension to this framework to test hybridized KEMs. Specifically, our extension allows for testing hybridized KEMs, where the KEM is constructed using a provably-secure combiner of the form k = F(k1, c) ⊕ F(k2, c) [16], where ⊕ is bitwise XOR, k1 and k2 are shared secrets from the component KEMs, c is the concatenated ciphertext, F is some function, and k is the resulting shared secret, by combining any KEM with any asymmetric cipher, such as RSA.
-
We test and demonstrate the methodological generality of our deep learning-based constructions by applying them also to other hybridized scenarios, such as cascade symmetric encryption <ref:2604. p. 2
06942#pg4>.Where the paper says it
• We test and demonstrate the methodological generality of our deep learningbased constructions by applying them also to other hybridized scenarios, such as cascade symmetric encryption [36]. • For the experiments above, we introduce rigorous statistical analysis, unlike existing deep learning-based results, through two-sided binomial hypothesis testing to determine whether the neural binary classifiers gain a nonnegligible advantage or not.
-
For the experiments above, we introduce rigorous statistical analysis, unlike existing deep learning-based results, through two-sided binomial hypothesis testing to determine whether the neural binary classifiers gain a nonnegligible advantage or not <ref:2604. p. 2
06942#pg4>.Where the paper says it
• We test and demonstrate the methodological generality of our deep learningbased constructions by applying them also to other hybridized scenarios, such as cascade symmetric encryption [36]. • For the experiments above, we introduce rigorous statistical analysis, unlike existing deep learning-based results, through two-sided binomial hypothesis testing to determine whether the neural binary classifiers gain a nonnegligible advantage or not.
How the training signal is built
For each KEM, a DNN is trained to distinguish between two classes of ciphertext, one class consisting of encryptions of uniformly random plaintexts, and another class consisting of encryptions of a fixed plaintext, in this case, all 0s <ref:2604.06942#pg4>. Each DNN is trained using stochastic gradient descent using the BCE as the loss function <ref:2604. p. 6 For
each KEM, a DNN is trained to distinguish between two classes of ciphertext,
one class consisting of encryptions of uniformly random plaintexts, and another
class consisting of encryptions of a fixed plaintext, in this case, all 0s. Each DNN
is trained using stochastic gradient descent using the BCE as the loss function.Where the paper says it
For KEMs using a combiner of the form k = F(k1, c) ⊕ F(k2, c), Algorithm 2 receives as input the concatenation of the ciphertext outputs of the two components <ref:2604. p. 7 2 The IND-CPA Testing Algorithm for KEMs using a combiner
of the form k = F(k1, c) ⊕ F(k2, c)
For scenarios in which a combiner of the form k = F(k1, c) ⊕ F(k2, c) (for example, the XOR combiner or the provably secure combiner) is used, we apply
Algorithm 2. In these experiments, the DNN receives as input the concatenation
of the ciphertext outputs of the two components. In these experiments, the DNN receives as input the concatenation
of the ciphertext outputs of the two components. The corresponding binary classification task distinguishes between (i) samples where the two ciphertext components decrypt/decapsulate to the same underlying value (the KEM shared
secret), and (ii) samples where they decrypt/decapsulate to two independent
values of the same length, generated independently of each other.Where the paper says it
Where the paper says it
Where the gains come from
The results in Table 4 show the robustness of the provably-secure combiner, as the neural classifier does not learn any patterns that allow it to classify the ciphertexts with any accuracy better than random guessing <ref:2604.06942#pg4>. Notably, even when an insecure component algorithm, such as plain RSA or plaintext, is used, the observed accuracies remain close to 50% under the considered DNN adversary model <ref:2604. p. 0 In our experiments on PQC algorithms, KEM combiners, and cascade
encryption, no algorithm or combination of algorithms demonstrates a
significant advantage (evaluated via two-sided binomial tests with significance level α = 0. 01), consistent with theoretical guarantees that
hybrids including at least one IND-CPA-secure component preserve indistinguishability, and with the absence of exploitable patterns under
the considered DNN adversary model. Each DNN
is trained using stochastic gradient descent using the BCE as the loss function. If the encryption scheme used is IND-CPA secure, the resulting DNN should not
be able to classify ciphertexts with an accuracy better than random guessing. Each DNN
is trained using stochastic gradient descent using the BCE as the loss function. If the encryption scheme used is IND-CPA secure, the resulting DNN should not
be able to classify ciphertexts with an accuracy better than random guessing.Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The paper concludes that no algorithm or combination of algorithms demonstrates a significant advantage (evaluated via two-sided binomial tests with significance level α = 0.01), consistent with theoretical guarantees that hybrids including at least one IND-CPA-secure component preserve indistinguishability, and with the absence of exploitable patterns under the considered DNN adversary model <ref:2604. p. 0 In our experiments on PQC algorithms, KEM combiners, and cascade
encryption, no algorithm or combination of algorithms demonstrates a
significant advantage (evaluated via two-sided binomial tests with significance level α = 0. 01), consistent with theoretical guarantees that
hybrids including at least one IND-CPA-secure component preserve indistinguishability, and with the absence of exploitable patterns under
the considered DNN adversary model. Finally, we describe our statistical evaluations
of the DNN performance. 1 The IND-CPA Testing Algorithm for Non-hybridized KEMs
and Cascade Encryption
We employ Algorithm 1 from [21] to model IND-CPA games using DNNs.Where the paper says it
Where the paper says it
Conclusion and Future Work
In this work, we applied the DNN IND-CPA classification framework to validate the IND-CPA property of the underlying PKE components used to construct three PQC algorithms, namely ML-KEM, HQC, and BIKE <ref:2604.06942#pg4>. Our results demonstrated that no classifier achieves an accuracy that is significantly different from random guessing, which is consistent with the algorithms being IND-CPA <ref:2604. p. 6 Each DNN
is trained using stochastic gradient descent using the BCE as the loss function. If the encryption scheme used is IND-CPA secure, the resulting DNN should not
be able to classify ciphertexts with an accuracy better than random guessing. The proposed neural-network
framework complements formal security analysis by providing a practical and
flexible, empirical tool to evaluate indistinguishability and identify potential
weaknesses in implementations and composed constructions. The main contributions of this work include:
• We use a DNN-based modeling of the IND-CPA game to empirically test the
IND-CPA properties of the underlying public key encryption (PKE) schemes
used in ML-KEM, BIKE, and HQC. The proposed neural-network
framework complements formal security analysis by providing a practical and
flexible, empirical tool to evaluate indistinguishability and identify potential
weaknesses in implementations and composed constructions. The main contributions of this work include:
• We use a DNN-based modeling of the IND-CPA game to empirically test the
IND-CPA properties of the underlying public key encryption (PKE) schemes
used in ML-KEM, BIKE, and HQC. If the encryption scheme used is IND-CPA secure, the resulting DNN should not
be able to classify ciphertexts with an accuracy better than random guessing. We remark that Algorithm 1 is not directly applicable to all KEMs, especially KEMs constructed with Fujisaki-Okamoto transform or variants thereof,
since Algorithm 1 requires selecting a plaintext, which is not possible for such
constructions.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The authors emphasize that the results in Table 4 show the robustness of the provably-secure combiner, as the neural classifier does not learn any patterns that allow it to classify the ciphertexts with any accuracy better than random guessing <ref:2604.06942#pg4>. The results in Table 5 and Table 6 empirically validate the ciphertext indistinguishability of cascade symmetric ciphers, as the neural classifier fails to classify the ciphertexts with any accuracy better than random guessing <ref:2604.06942#pg4>. The results shown in Tables 5 and 6 empirically validate the ciphertext indistinguishability of cascade symmetric ciphers, as the neural classifier fails to classify the ciphertexts with any accuracy better than random guessing <ref:2604. p. 6 Each DNN
is trained using stochastic gradient descent using the BCE as the loss function. If the encryption scheme used is IND-CPA secure, the resulting DNN should not
be able to classify ciphertexts with an accuracy better than random guessing.Where the paper says it
The paper presents rigorous statistical analysis to validate that the accuracy of our trained models is statistically indistinguishable from that of random guessing <ref:2604.06942#pg4>. The results demonstrate the versatility of using a DNN classifier to validate the IND-CPA security of both individual algorithms and composite constructions, including hybrid KEMs and cascade encryption <ref:2604. p. 2 1 Main Contributions
This paper extends a DNN-based approach used for the IND-CPA testing, originally proposed in [21], to demonstrate its applications in the testing of PQC
KEMs, and further extends it to test hybrid KEMs. The proposed neural-network
framework complements formal security analysis by providing a practical and
flexible, empirical tool to evaluate indistinguishability and identify potential
weaknesses in implementations and composed constructions. In our experiments on PQC algorithms, KEM combiners, and cascade
encryption, no algorithm or combination of algorithms demonstrates a
significant advantage (evaluated via two-sided binomial tests with significance level α = 0. 01), consistent with theoretical guarantees that
hybrids including at least one IND-CPA-secure component preserve indistinguishability, and with the absence of exploitable patterns under
the considered DNN adversary model.Where the paper says it
Where the paper says it
The authors note that the perfect accuracy is expected for deterministic RSA, while applying the neural classifier to RSA-OAEP achieved performance that is statistically indistinguishable from random guessing, which is consistent with RSA-OAEP’s proven IND-CPA security <ref:2604.06942#pg4>. This consistency confirms that the DNN approach successfully validates established security properties for PQC schemes <ref:2604. p. 2 The proposed neural-network
framework complements formal security analysis by providing a practical and
flexible, empirical tool to evaluate indistinguishability and identify potential
weaknesses in implementations and composed constructions. The main contributions of this work include:
• We use a DNN-based modeling of the IND-CPA game to empirically test the
IND-CPA properties of the underlying public key encryption (PKE) schemes
used in ML-KEM, BIKE, and HQC. 1 Main Contributions
This paper extends a DNN-based approach used for the IND-CPA testing, originally proposed in [21], to demonstrate its applications in the testing of PQC
KEMs, and further extends it to test hybrid KEMs. The proposed neural-network
framework complements formal security analysis by providing a practical and
flexible, empirical tool to evaluate indistinguishability and identify potential
weaknesses in implementations and composed constructions.Where the paper says it
Where the paper says it
The research provides an adaptive, practical, and versatile empirical estimator for indistinguishability in more general IND-CPA settings, allowing data-driven validation of implementations and compositions and complementing the analytical security analysis <ref:2604.06942#pg4>. This illustrates the potential of using deep learning as an adaptive, practical, and versatile empirical estimator for indistinguishability in more general IND-CPA settings <ref:2604. p. 0 These illustrate the potential of
using deep learning as an adaptive, practical, and versatile empirical
estimator for indistinguishability in more general IND-CPA settings, allowing data-driven validation of implementations and compositions and
complementing the analytical security analysis. Keywords: Deep Learning for Post-Quantum Cryptography Transition
· Hybrid
encryption
· IND-CPA
· Combiners
arXiv:2604. This led the National Institute of Standards and Technology
(NIST) [26] to initiate a standardization process for post-quantum cryptography
(PQC) algorithms that, in 2024, resulted in the standardization of ML-KEM [28]
for key exchange; SLH-DSA [29] and ML-DSA [27] for digital signatures. Due to
long replacement times for cryptographic primitives and "harvest now, decrypt
later" attacks [24], the transition to PQC is currently of high priority [14]. Such
hybrid methods can, in principle, also be used to combine multiple different PQC
algorithms without any classical component, to provide similar risk mitigation
in scenarios where the classical algorithms have been compromised. We remark that the connections between cryptography and machine learning methods were mentioned as early as 1991 by Rivest [32], drawing a parallel
between machine learning and cryptanalysis as two cases in which one tries to
learn an unknown function. Adaptive IND-CPA Testing of PQC KEMs 7
3 Deep Learning, KEM, and Cascade Encryption Setups
for IND-CPA Testing
In this section, we introduce the DNN models for testing IND-CPA and then apply them to the PKE schemes used to construct PQC KEMs, such as ML-KEM,
BIKE, and HQC. We then present our extended algorithm for testing the INDCPA of hybrid methods using combiners of the form k = F(k1, c) ⊕ F(k2, c), as
described in Definition 8.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive IND-CPA Testing for PQC KEMs: The system can empirically validate that no algorithm or combination of algorithms demonstrates a significant advantage against deep learning adversaries, which helps in data-driven validation of implementations and compositions.
-
Hybrid KEM Distinguishability Testing: The system can test hybridized KEMs using the combiner form k = F(k1, c) ⊕ F(k2, c), allowing it to assess security inheritance from stronger components under a specific DNN adversary model.
-
Cascade Encryption Security Assessment: The system can evaluate the ciphertext indistinguishability of cascade symmetric encryption by applying Algorithm 1 to combinations of AES and ChaCha20, showing that
the classifier achieved accuracy indistinguishable from random guessing whenever at least one component algorithm is IND-CPA.
Abstract
We model indistinguishability under chosen-plaintext attack (IND-CPA) games as binary classification tasks for deep neural networks (DNNs). We apply this methodology to the public-key encryption (PKE) schemes used to construct PQC key encapsulation mechanisms (KEMs) such as ML-KEM and HQC, and to cascaded symmetric algorithms, demonstrating methodological generality. Moreover, we present a novel extension of this approach to empirical distinguishability testing of hybrid KEMs. For IND-CPA-secure schemes and their combinations, no DNN achieves accuracy significantly different from random guessing (measured via Bonferroni-corrected binomial tests), consistent with theoretical security guarantees. For insecure schemes, such as plain RSA, AES-CTR and AES-CBC with IV or counter reuse, or RC4, the DNNs classify the ciphertexts better than random guessing. Thus, this framework is shown to serve as a practical, scheme-agnostic tool for flagging distinguishable ciphertexts in both individual and composed constructions.
Sources
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data