acceptodds
Withdrawn from ICLR 2027

LMND: Language-Model Neural Distinguishers via Conditional Generative Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Block ciphers are a cornerstone of modern secure communication. Differential cryptanalysis assesses their security by distinguishing ciphertext pairs generated under a fixed input difference from random ones. Neural differential methods replace hand-crafted differential statistics with trained deep neural networks, learning directly from data the weak residual dependencies that are difficult to characterize explicitly, and thereby extending effective analysis to more rounds. Existing CNN-based neural distinguishers, typically cast this task as binary classification: each ciphertext pair receives only one sample-level label, yielding coarse supervisory granularity, while the sigmoid output lacks an explicit probabilistic reference, making it difficult to endow the score with statistical meaning or to accumulate evidence across samples. To address these issues, we propose LMND, a neural distinguisher built on conditional generative modeling. Conditioned on one ciphertext, LMND autoregressively predicts the eight nibble tokens of the other, providing position-level supervision for every real pair. Exploiting the fact that fixed-key encryption is a permutation, it takes the uniform distribution as an analytic reference and converts the conditional likelihood into a distinguishing score that requires no negative training samples, no threshold fitting, and accumulates across pairs, with shared-weight bidirectional scoring enforcing invariance under ciphertext exchange. On 58-round Speck32/64, LMND attains the highest single-pair accuracy among all evaluated models; in a 9-round subkey-ranking experiment, it reduces the mean rank from to and raises the rank-0 success rate from to relative to Gohr's baseline.

Primary area:
generative models
Keywords:
differential cryptanalysis, neural distinguishers, conditional generative modeling, block ciphers, Speck32/64

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6xN9OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper uses conditional generative modeling in the field of block cipher. In comparison, existing works use CNN which is a classification method. However, it seems that the method (transformers) is lack of novalty, and does not provide many insights for the ML community. I suggest the authors submit this manuscript to a journal of security instead of ICLR. [I check the CNN baseline, which is also not published in the ML conferences. ]

Critical strengths

This paper goes beyond existing classification techniques to generative techniques in the field of block cipher, and achieves better performances.

Critical weaknesses

I am not an expert in this field, and I can only provide suggestions from a ML view. The main concerns here is the technique itself is not interesting. It seems that this paper only applies an existing technique to a different task and achieves a better performance. This seems not meet the ICLR standard.

Questions for authors

I suggest the authors provide evidence on how this paper provides insights in the ML community.

Reviewer_mDpZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores how to conduct a neural distinguisher-based attack on the SPECK cipher. The authors use a neural distinguisher that is built on conditional generative modeling. Then, conditioned on one ciphertext, the distinguisher autoregressively predicts the eight nibble tokens of the other, providing position-level supervision for every real pair. The experiments are done on SPECK 32/64 for 5 to 8 rounds. Moreover, the authors consider 9 round subkey ranking experiment. The results indicate improved performance compared to Gohr's baseline.

Critical strengths
  • Interesting new approach for neural distinguisher-based attacks.
Critical weaknesses
  • Limited experiments and comparison with related work.

  • The explainability of the proposed approach is not discussed.

Questions for authors

The paper introduces as one of the key challenges "Second, training depends on random negative samples.". I do not undestand why is this key challenge? What exactly is the problem?

The model uses d_{model} = 128, eight attention heads, an FFN width of 512, and 1,857,664 parameters. -> why these settings? How are those obtained?

Regarding the experiments, I have several questions. First, while the paper states it considers SPECK32/64, I cannot understand, per the tables, for which version the results are? Since nothing is stated, does it mean the attack performs exactly the same for both SPECK32 and SPECK64? If we look at recent SoK on neural distinguishers, it seems those two cases generally have different results? See https://cic.iacr.org/p/3/2/21/pdf Moreover, while I understand the comparison with Gohr's work since it is the first one on the topic, I would say that the paper does not provide sufficient related work comparison. How did the authors decide to compare with Hou et al. additionally? In general, how do your results compare with wider related work discussed in https://cic.iacr.org/p/3/2/21/pdf ? Next, attacks on SPECK can go for more rounds than investigated here. Why not look at the 10-round case as well?

Stronger single-pair distinguishing can improve key recovery by helping to separate correct and incorrect subkey candidates -> please elaborate on this claim

Section 4.5 discusses interesting complementarity between CNNs and Transformers. However, besides the fact that the authors report there is complementarity, I do not find so much useful information. Why does complementarity happen? What mechanisms are responsible for it?

Finally, one of the main challenges for neural distinguishers is their black-box character, where cryptanalysis requires understanding of the attack. What is the situation with the LMND? Are the results more explainable/interpretable than with classical CNNs?

Reviewer_nmEsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LMND, a conditional encoder–decoder Transformer that predicts one ciphertext as eight autoregressive nibble tokens conditioned on the other. The learned likelihood is compared against a uniform reference, with bidirectional averaging for distinguishing and score accumulation for subkey ranking. Experiments cover 5–8-round SPECK32/64 distinguishing and 9-round last-round subkey ranking, alongside token-order/layout ablations and CNN-complementarity analysis. Compared with Gohr’s baseline, the reported mean subkey rank decreases from 52.1 to 23.7, and rank-0 success increases from 35.8% to 41.6%.

Critical strengths
  1. The conditional-density formulation provides a coherent positive-only learning objective and an analytically justified uniform reference under the stated sampling assumptions. This offers an interpretable alternative to discriminative neural distinguishers.
  2. The evaluation extends beyond classification accuracy: 1,000 subkey-ranking trials, rank quantiles, and success rates provide evidence of downstream cryptanalytic utility. The token-order and conditioning-layout ablations also identify useful design choices.
Critical weaknesses
  1. The motivation does not establish that binary supervision or statistically uninterpretable scores are the relevant bottlenecks: posterior odds from a Bayes-optimal balanced classifier recover the density ratio, and Gohr already accumulates log-odds for key ranking. The paper should distinguish its contribution from standard conditional density estimation and explain which cipher-specific learning limitation the generative objective resolves. Gohr’s implementation: https://github.com/agohr/deep_speck/blob/master/key_rank.py
  2. Section 4.4 tests token order and conditioning layout, but does not isolate the claimed benefits of position-level supervision, positive-only training, or bidirectional scoring. A matched Transformer with a binary objective and one-direction versus two-direction scoring controls are needed to separate the proposed objective from architecture, representation, and model capacity.
  3. The empirical coverage is narrow: all results concern SPECK32/64, with a single main input difference and only two single-pair CNN baselines. Another cipher structure or block size, additional differences, and stronger relevant baselines would help establish whether the improvement extends beyond this particular setting.
  4. Table 1 reports variation over five test batches rather than independent training runs, so its small standard deviations do not establish training robustness. Independent seeds and paired uncertainty estimates for the attack gains are needed, while the accuracy differences between Tables 1 and 2—such as 0.933 versus 0.9296 at five rounds—require clarification.
  5. The practical advantage remains insufficiently quantified: subkey ranking uses only 64 pairs and compares with Gohr and DDT, without reporting inference time or matched training costs. Success-versus-data and runtime comparisons, including ND1, are needed to establish whether the ranking gains justify the computational cost; the current explanations involving finer supervision and accumulation remain largely unverified.
Questions for authors
  1. Do the five evaluations in Table 1 share one trained checkpoint, and which checkpoint or sampling differences explain the discrepancies between Tables 1 and 2?
  2. Were the baselines matched for positive-pair exposure, optimization steps, curriculum, EMA, and model selection, and which saved LMND weights were evaluated?
  3. Can the authors compare generative and binary objectives using the same Transformer and representation, and separately quantify the benefit of bidirectional scoring?
  4. How do LMND and ND1 compare on identical key-ranking instances across several pair budgets, including inference time and paired confidence intervals? If the author addresses my concerns, I would be willing to raise my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.