In recent years, the integration of deep learning with differential cryptanalysis has led to differential neural cryptanalysis, enabling efficient data-driven security evaluation of modern cryptographic algorithms. Compared to traditional cryptanalysis, differential neural cryptanalysis enhances the efficiency and automation of the analysis by training neural networks to automatically extract statistical features from ciphertext pairs. As research advances, neural distinguisher construction faces challenges due to the absence of a unified framework capable of cross-algorithm generalization and feature optimization. There is no systematic way to build a framework from adapted data formats and network architectures, which limits their scalability across diverse ciphers and their suitability for combining different cryptanalysis methods. Besides, neural network training is data-driven, there is an urgent need to combine cryptographic theory with data analysis methods to systematically evaluate the quality of differentially generated datasets. To address these gaps, this paper proposes a novel framework for constructing related-key neural differential distinguishers that integrates three core innovations: (1) multi-ciphertext multi-difference formats to enhance dataset diversity and feature coverage, (2) structural filtering for prioritizing high-probability differential paths aligned with cryptographic architectures, and (3) Deep Residual Shrinkage Network (DRSN) with adaptive thresholding to suppress noise and amplify critical differential features. By applying this framework to two standardized algorithms DES and PRESENT, our results demonstrate data-driven breakthroughs in standard ciphers. For DES, the framework achieves an 8-round related-key neural distinguisher and improves 6/7-round distinguisher accuracy by over 40%. For PRESENT, we construct the first 9-round related-key neural distinguisher, which outperforms existing neural distinguishers in both round coverage and accuracy. Additionally, we systematically analyze dataset quality using kernel principal component analysis (KPCA) and K-means clustering, revealing a strong correlation between clustering compactness and distinguisher performance. Furthermore, we propose a validation algorithm to verify differential combinations with cryptographic advantages from a machine learning perspective, identifying ‘good’ plaintext-key differential combinations. We apply this approach to the SIMECK algorithm, demonstrating its broad applicability.

Encrypted traffic classification is crucial for critical network management tasks such as traffic type identification, resource allocation, and risk mitigation, especially given that encrypted traffic has become the dominant form of modern network communication. However, existing classification methods are typically confined to single-level feature extraction, failing to capture the multi-granularity information inherent in traffic and thus limiting their ability to characterize complex encrypted traffic patterns. To address this issue, this paper proposes BPF-GNN, a hierarchical graph feature extraction model for encrypted traffic classification. The model enables multi-granularity feature learning by constructing a three-tier graph structure (Byte-, Packet-, and Flow-level). It sequentially extracts discriminative information inherent in each granularity level and accumulates multi-dimensional traffic characteristics, significantly improving the classification accuracy of encrypted traffic. Experiments on the ISCX-VPN2016, ISCX-Tor2016, USTC-TFC2016, and MIRAGE-2024 datasets demonstrate that BPF-GNN outperforms existing methods, validating the effectiveness and superiority of the proposed hierarchical multi-granularity feature extraction approach.

Abstract In CRYPTO 2019, Gohr built a bridge between machine learning and differential cryptanalysis, which show that machine learning-aided methods have advantages over classical differential cryptanalysis. Yet, for linear cryptanalysis, there is lack of effective works showing that machine learning-aided cryptanalysis can reach the benchmark of traditional counterparts and also lack of an effective universal framework using machine learning to assist linear cryptanalysis. In this paper, we mainly focus on machine learning-aided linear cryptanalysis and application to Des. First, we propose a machine learning-aided model to distinguish different Bernoulli distributions and demonstrate the validity of the model through experiments and theoretical analysis. Based on the model, we propose a new machine learning-aided linear cryptanalysis framework, which can be applied to one bit and multiple bits key-recovery attacks. As applications, we perform one bit attacks on 3-, 4-, 5-, 6-round Des and multiple bits attack on 8-round Des. Compared with the previous works about machine learning-aided linear cryptanalysis, the results improve the success rate and the complexity. Most importantly, more rounds are covered in our work. Besides, the work indicates that machine learning-aided cryptanalysis can achieve the same or marginally better performance than classical methods.

At the Annual International Cryptology Conference in 2019, Gohr introduced a deep learning based cryptanalysis technique applicable to the reduced-round lightweight block ciphers with a short block of SPECK32/64. One significant challenge left unstudied by Gohr’s work is the implementation of key recovery attacks on large-state block ciphers based on deep learning. The purpose of this paper is to present an improved deep learning based framework for recovering keys for large-state block ciphers. First, we propose a key bit sensitivity test (KBST) based on deep learning to divide the key space objectively. Second, we propose a new method for constructing neural distinguisher combinations to improve a deep learning based key recovery framework for large-state block ciphers and demonstrate its rationality and effectiveness from the perspective of cryptanalysis. Under the improved key recovery framework, we train an efficient neural distinguisher combination for each large-state member of SIMON and SPECK and finally carry out a practical key recovery attack on the large-state members of SIMON and SPECK. Furthermore, we propose that the 13-round SIMON64 attack is the most effective approach for practical key recovery to date. Noteworthly, this is the first attempt to propose deep learning based practical key recovery attacks on 18-round SIMON128, 19-round SIMON128, 14-round SIMON96, and 14-round SIMON64. Additionally, we enhance the outcomes of the practical key recovery attack on SPECK large-state members, which amplifies the success rate of the key recovery attack in comparison to existing results.

Abstract At CRYPTO’19, Gohr built a bridge between deep learning and cryptanalysis. Based on deep neural networks, he trained neural distinguishers of SPECK32/64. Besides, with the help of neural distinguishers, he attacked 11-round SPECK32/64 using Bayesian optimization. Compared with the traditional attack, its complexity was reduced. Although his work opened a new direction of machine learning aided cryptanalysis, there are still two research gaps that researchers are eager to fill in. (i) Can the attack using neural distinguishers be used to other block ciphers? (ii) Are there effective key recovery attacks on large-size block ciphers adopting neural distinguishers? In this paper, our core target is to propose an effective neural-aided key recovery policy to attack large-size block ciphers. For large-size block ciphers, it costs too much time in pre-computation, especially in wrong key response profile, which is the main reason why there are almost no neural aided attacks on large-size block ciphers. Fortunately, we find that there is a fatal flaw in the wrong key profile. In the some experiments of SIMON32/64 and SIMON48/96, there is a regular of change in response profiles, which implies that we can use partial response instead of the complete response. Based on this, we propose a generic key recovery attack scheme which can attack large-size block ciphers. As an application, we perform a key recovery attack on 13-round SIMON64/128, which is the first practical attack using neural distinguishers to large-size ciphers. In addition, we also attack 13-round SIMON32/64 and SIMON48/96, which also shows that the neural distinguishers can be used to other block ciphers.

Abstract In CRYPTO 2019, Gohr opens up a new direction for cryptanalysis. He successfully applied deep learning to differential cryptanalysis against the NSA block cipher SPECK32/64, achieving higher accuracy than traditional differential distinguishers. Until now, one of the mainstream research directions is increasing the training sample size and utilizing different neural networks to improve the accuracy of neural distinguishers. This conversion mindset may lead to a huge number of parameters, heavy computing load, and a large number of memory in the distinguishers training process. However, in the practical application of cryptanalysis, the applicability of the attacks method in a resource-constrained environment is very important. Therefore, we focus on the cost optimization and aim to reduce network parameters for differential neural cryptanalysis.In this paper, we propose two cost-optimized neural distinguisher improvement methods from the aspect of data format and network structure, respectively. Firstly, we obtain a partial output difference neural distinguisher using only 4-bits training data format which is constructed with a new advantage bits search algorithm based on two key improvement conditions. In addition, we perform an interpretability analysis of the new neural distinguishers whose results are mainly reflected in the relationship between the neural distinguishers, truncated differential, and advantage bits. Secondly, we replace the traditional convolution with the depthwise separable convolution to reduce the training cost without affecting the accuracy as much as possible. Overall, the number of training parameters can be reduced by less than 50% by using our new network structure for training neural distinguishers. Finally, we apply the network structure to the partial output difference neural distinguishers. The combinatorial approach have led to a further reduction in the number of parameters (approximately 30% of Gohr’s distinguishers for SPECK).