Improved differential-neural cryptanalysis for round-reduced SIMECK32/64

Liu Zhang, Jinyu Lu, Zilong Wang, Chao Li

TL;DR

This paper presents an improved neural network architecture for cryptanalysis of the SIMECK32/64 lightweight block cipher. By enhancing the neural network's ability to capture differential characteristics, the authors achieve practical key recovery attacks on up to 17 rounds of SIMECK32/64, extending previous state-of-the-art by two rounds.

What Problem Does This Paper Address?

Lightweight block ciphers are crucial for data confidentiality in resource-constrained environments. However, their security is paramount, and efficient cryptanalysis techniques are essential. Differential cryptanalysis, a powerful method, has been augmented by machine learning, particularly deep learning, to improve its effectiveness. Previous work by Gohr et al. introduced differential-neural cryptanalysis (ND) for SIMECK32/64, achieving practical attacks on 11- and 12-round versions. However, extending these attacks to more rounds remains a challenge due to limitations in the accuracy of neural distinguishers and the complexity of key recovery. This paper aims to improve the accuracy of neural distinguishers for SIMECK32/64 and leverage these improvements to mount practical key recovery attacks on a greater number of rounds.

Methods

The core of this work lies in improving the neural network architecture for differential-neural cryptanalysis. The authors adapt the "Inception" neural network, previously used for SIMON and SPECK, to better fit the round function of SIMECK. This involves:

  • Modified Inception Architecture: The Inception module is adjusted to better capture the specific cyclic rotation and XOR operations within the SIMECK round function.
  • Multiple Ciphertext Pairs: The neural network is trained using multiple ciphertext pairs as input for each sample. This approach helps in capturing more nuanced differential information and improving the accuracy of the neural distinguisher.
  • Staged Training: A staged training method is employed to progressively train the neural network, starting with recognizing differences in fewer rounds and gradually increasing to the target number of rounds. This helps in achieving higher accuracy.
  • Wrong Key Response Profile (WKRP) Analysis: The authors conduct a detailed analysis of the wrong key response profile. This analysis reveals which bits of the subkey have a significant impact on the distinguisher's score, allowing for optimization of the Bayesian key search algorithm. Specifically, they identify that the 14th and 15th bits have a larger impact, while the 12th and 13th bits have less influence, enabling faster key recovery.
  • Enhanced Key Recovery: The improved neural distinguishers, combined with insights from the WKRP analysis and optimized Bayesian key search, are used to mount practical key recovery attacks.

Datasets

The paper utilizes datasets generated for training and evaluating the neural distinguishers.

  • Generated Datasets: The datasets consist of pairs of plaintexts with a specific input difference and their corresponding ciphertexts. For training, these pairs are generated with uniformly distributed keys and specific input differences (0x0000, 0x0040). The generation process involves encrypting plaintext pairs for a certain number of rounds, and for negative samples, one of the plaintexts is replaced with a random one before encryption.
  • No New Benchmarks Introduced: The paper does not introduce new public datasets but rather generates specific datasets for their experiments based on the SIMECK32/64 cipher.

Key Insights and Ideas

  • Architecture Matters: Adapting neural network architectures to the specific round function of a cipher is crucial for improving cryptanalytic performance. The modified Inception network proved more effective for SIMECK than generic architectures.
  • More Data, More Insight: Using multiple ciphertext pairs per sample significantly boosts the accuracy of neural distinguishers, especially for distinguishing between correct and incorrect key guesses.
  • WKRP is Key to Optimization: Understanding the wrong key response profile is vital for accelerating key recovery. Identifying "insensitive" key bits allows for more efficient search strategies.
  • Synergy of Classical and Neural: Combining classical differential techniques (CDs) with neural distinguishers (NDs) remains a powerful strategy for extending cryptanalytic attacks to more rounds.

Key Results

  • Improved Accuracy: The proposed Inception-based neural network achieves improved accuracy for (9-12)-round neural distinguishers for SIMECK32/64 compared to previous methods.
  • Practical Attacks: The authors successfully demonstrate practical key recovery attacks on up to 17 rounds of SIMECK32/64.
  • High Success Rates: The 15-round and 16-round attacks achieve an almost 100% success rate.
  • Efficiency Gains: The analysis of the wrong key response profile leads to an optimized key recovery process, reducing the overall complexity. For instance, the 15-round attack has a success rate of 99.17% and a time complexity of approximately 2^235.309. The 16-round attack achieves 100% success rate with a time complexity of approximately 2^238.189. The 17-round attack has a success rate of 30% with a time complexity of approximately 2^245.037.

Limitations and Future Work

  • Data Complexity: While the attacks are practical, the data complexity can still be significant for higher rounds. The paper notes that the theoretical data complexity is lower than the actual in experiments due to high accuracy.
  • Computational Resources: The experiments require substantial computational resources, including multiple GPUs.
  • Future Work: The authors suggest that further exploration of neutral bits and the wrong key response profile could lead to attacks on even more rounds. Additionally, investigating other neural network architectures and training strategies could yield further improvements. The paper also hints at the possibility of extending these techniques to other lightweight ciphers.

Summary updated

Abstract

In CRYPTO 2019, Gohr presented differential-neural cryptanalysis by building the differential distinguisher with a neural network, achieving practical 11-, and 12-round key recovery attack for Speck32/64. Inspired by this framework, we develop the Inception neural network that is compatible with the round function of Simeck to improve the accuracy of the neural distinguishers, thus improving the accuracy of (9-12)-round neural distinguishers for Simeck32/64. To provide solid baselines for neural distinguishers, we compute the full distribution of differences induced by one specific input difference up to 13-round Simeck32/64. Moreover, the performance of the DDT-based distinguishers in multiple ciphertext pairs is evaluated. Compared with the DDT-based distinguishers, the 9-, and 10-round neural distinguishers achieve better accuracy. Also, an in-depth analysis of the wrong key response profile revealed that the 12-th and 13-th bits of the subkey have little effect on the score of the neural distinguisher, thereby accelerating key recovery attacks. Finally, an enhanced 15-round and the first practical 16-, and 17-round attacks are implemented for Simeck32/64, and the success rate of both the 15-, and 16-round attacks is almost 100%.
BibTeX
@article{zhang2023improved,
  author = {Liu Zhang and Jinyu Lu and Zilong Wang and Chao Li},
  title = {{Improved differential-neural cryptanalysis for round-reduced SIMECK32/64}},
  journal = {Frontiers of Computer Science},
  year = {2023},
  doi = {10.1007/s11704-023-3261-z},
}
Figure