Abstract
This paper presents enhanced neural distinguishers for the SIMON and SIMECK block ciphers by integrating cryptographic domain knowledge into deep learning models. By utilizing a Squeeze-and-Excitation Residual Network (SE-ResNet) architecture and domain-informed data formatting, the authors achieve higher accuracy and extend the number of rounds covered compared to previous state-of-the-art methods. Furthermore, the study introduces the first related-key neural distinguishers for these ciphers, demonstrating significant improvements in cryptanalytic capabilities and revealing structural vulnerabilities in SIMECK’s key-expansion algorithm.
Introduction
In modern cryptanalysis, a "distinguisher" is a fundamental tool used to identify non-random behavior in a cipher's output. Proving that a cipher can be distinguished from random noise often serves as a precursor to full key-recovery attacks. Traditionally, this has relied on mathematical techniques like differential cryptanalysis. Recently, deep learning has emerged as a powerful alternative, capable of automatically learning complex patterns in ciphertext. However, early "black-box" neural distinguishers often ignored the internal structure of the ciphers they targeted. This research bridges that gap by injecting domain-specific cryptographic knowledge into the training process to improve the performance of neural distinguishers for the SIMON and SIMECK families of lightweight block ciphers.
Methodology
SE-ResNet Architecture
The researchers employed a sophisticated deep learning architecture known as a Squeeze-and-Excitation Residual Network (SE-ResNet). This model enhances standard residual networks by adding SE blocks, which allow the network to perform dynamic channel-wise feature recalibration. This architecture is particularly effective at capturing the subtle, non-linear dependencies present in cryptographic data.
Domain-Informed Data Formatting
A key innovation of this work is the move away from raw ciphertext pairs. Instead, the network is provided with a complex tuple of features derived from cryptographic theory, including partial differences from previous rounds: $(\Delta^r_L, \Delta^r_R, C_l, C_r, C'_l, C'_r, \Delta^{r-1}_R, p\Delta^{r-2}_R)$. The inclusion of the $p\Delta^{r-2}_R$ term—a partial difference from two rounds prior—was found to be a "secret sauce" that significantly boosted the network's ability to distinguish cipher output from noise.
Staged Training and Input Search
To overcome the vanishing signal in high-round distinguishers, the authors utilized staged training, where weights from a model trained on fewer rounds serve as the initialization for a more complex task. Additionally, they performed a systematic search of thousands of input differences with low Hamming weights to identify the optimal starting difference $(\Delta_p)$ for maximizing accuracy.
Experimental Results
The proposed methods were tested against SIMON32/64, SIMON64/128, SIMECK32/64, and SIMECK64/128. The results demonstrate clear improvements over previous state-of-the-art neural distinguishers, such as those proposed by Gohr.
As shown in the table above, the researchers successfully extended the range of neural distinguishers. For instance, an 11-round distinguisher for SIMON32/64 achieved an accuracy of 0.5609, surpassing previous benchmarks. More significantly, the study presents the first-ever related-key neural distinguishers (RKND) for these ciphers, reaching up to 13 rounds for SIMON32/64 and 15 rounds for SIMECK32/64 with high accuracy. The results also indicate that SIMON’s key-expansion algorithm provides better resistance to these attacks than SIMECK’s, despite the similarities in their round functions.
Conclusion
By combining deep learning with traditional cryptographic insights, this research significantly advances the state of neural cryptanalysis. The use of domain-informed features and advanced architectures like SE-ResNet allows for more powerful distinguishers that can reach more rounds than pure data-driven approaches. While these distinguishers currently cover only a fraction of the total rounds in full ciphers, they provide a strong foundation for hybrid attacks. Future work will explore more aggressive feature engineering and alternative neural architectures to further push the boundaries of automated cryptanalysis.
Summary updated