Introduction
Since the seminal work of Aron Gohr in 2019, the field of neural differential cryptanalysis has undergone rapid expansion. Traditional differential cryptanalysis—the study of how input differences propagate through a cipher to leak secret keys—has long been a manual, mathematically intensive process requiring deep domain expertise. Gohr’s breakthrough demonstrated that deep learning could automate the creation of "neural distinguishers," which in some cases outperform state-of-the-art manual techniques, particularly on lightweight block ciphers like SPECK32.
This paper provides a comprehensive "Systematization of Knowledge" (SoK) for the field over the last five years. It addresses the "comparability crisis" caused by fragmented research methodologies by proposing a unified taxonomy, comparing the performance of various AI-driven distinguishers against 24 different cryptographic primitives, and highlighting the shift toward automated, explainable, and hybrid cryptanalytic attacks.
The Neural Distinguisher
The fundamental building block of this research area is the neural distinguisher ($\mathcal{ND}$). In a standard differential attack, a cryptanalyst seeks a differential characteristic that occurs with a probability significantly higher than random. A neural distinguisher automates this by learning to identify these biases directly from data.
As illustrated in Figure 1, the model is typically trained on ciphertext pairs $(C_0, C_1)$. One set of pairs is generated from a fixed input difference $\delta$ ($P_0 \oplus P_1 = \delta$), while the other set is generated from random differences. The $\mathcal{ND}$ learns to output a probability score: a value near 1 indicates the presence of the target difference, while a value near 0 indicates random noise.
Methodology and Key Recovery Pipelines
While a distinguisher can identify whether a cipher is behaving non-randomly, the ultimate goal of cryptanalysis is key recovery. This requires integrating the $\mathcal{ND}$ into a larger "Key Recovery Pipeline."
Figure 2 depicts the adversarial model for such an attack. The adversary interacts with an encryption oracle to obtain ciphertext pairs with a known input difference. They then guess a portion of the last round's key ($rk$) and partially decrypt the ciphertexts. The resulting intermediate values are fed into the neural distinguisher $\mathcal{ND}_r$. If the guess is correct, the $\mathcal{ND}$ will detect the expected differential bias ($p \approx 1$). If incorrect, the data will appear random ($p \ll 1$). Modern pipelines enhance this process using Bayesian search and "Neutral Bits" to extend the attack across more rounds than the distinguisher was originally trained for.
Taxonomy and Systematic Review
To organize the diverse body of literature, the authors employ a refined $n-m-T-E$ classification scheme:
- $n$: The number of ciphertexts per sample (e.g., single pairs vs. multiple pairs).
- $m$: The number of distinct input differences considered by the model.
- $T$: The type of feature engineering applied (e.g., using raw ciphertext, XOR differences, or bit-level representations).
- $E$: The type of experiment, ranging from standard detection to real-world ciphertext masking scenarios.
This framework covers a variety of architectures, from the foundational CNN-based $ND_{Gohr}$ to attention-based models like SE-ResNet and cipher-agnostic designs like DBitNet.
Key Insights and Results
The systematic review reveals several critical findings:
- Explainability and Differential-Linear Features: Using XAI tools like LIME and SHAP, researchers discovered that neural networks often learn "differential-linear" features—complex combinations of bitwise differences and linear approximations that are difficult for humans to model manually.
- The Importance of Feature Engineering: Data formatting often outweighs model complexity. For example, researchers found they could achieve near-identical accuracy by truncating 32-bit ciphertexts down to just the 8 most informative bits, suggesting that cryptographic leakage is highly concentrated.
- Scaling Limits: Unlike other AI domains, neural cryptanalysis does not follow a simple "bigger is better" scaling law. Deeper networks often overfit specific training keys and fail to generalize to the "random key" setting necessary for practical attacks.
- Benchmark Performance: Hybrid approaches have successfully extended key recovery on SPECK32/64 to 14 and 15 rounds. Automated pipelines like DBitNet have generated distinguishers for up to 13 rounds of SIMON32/64 without requiring any prior cryptanalytic knowledge.
Conclusion and Future Work
Despite significant progress, the field faces a "state-size barrier." Scaling these attacks to 128-bit ciphers like full AES remains a massive computational hurdle due to the curse of dimensionality. Furthermore, the authors highlight a need for better reproducibility and standardized benchmarking. They conclude by proposing a unified "leaderboard" for neural distinguishers to ensure that future research remains scientifically rigorous and comparable across the community.
Summary updated