Recently, differential-neural cryptanalysis, which combines deep learning with differential cryptanalysis, has gained attention as a powerful and practical cryptanalysis method. In this approach, a neural distinguisher is trained to distinguish the target cipher with specific differential properties from other ciphers, and its statistical analysis is used to guess the key. This approach offers the advantage of enabling deep learning to analyze cryptographic properties, traditionally demanding substantial time and expertise. Nevertheless, the black-box nature of deep learning models poses challenges for cryptanalysts in comprehending the construction of the differential dataset and the neural distinguisher. In particular, since the differential dataset serves as the foundation for generating the neural distinguisher, there is a pressing need for an effective methodology to construct a high-quality differential dataset. However, the existing approaches for the generation of differential dataset is based on the well-known differential characteristic or exploring through the training neural distinguisher. In this paper, we analyzed the intrinsic property of differential dataset and proposed a novel approach to construct a good differential dataset for differential-neural cryptanalysis. First, we conducted an analysis to find the difference between well-trainable differential datasets and other differential datasets using Principal Component Analysis (PCA) and K-means clustering. As a result, the well-trainable differential dataset has a property that there are some principal components who can well express the variance of the whole dataset, the data construct clusters in the principal components-based reduced dimension. Building upon our analysis results, we proposed the exploring algorithm for the generation of well-trainable differential dataset. This proposed algorithm assesses the input differences within the differential dataset by identifying significant principal components. If such components are found, a cluster evaluation is performed on the differential dataset using the silhouette score, which quantifies the quality of cluster formation. In our experiments, the proposed algorithm successfully identified favorable input differences, leading to improved accuracy in neural distinguisher training for SPECK and SIMON. Compared to the performance of the existing Gohr's neural input difference algorithm, our proposed algorithm was more effective in finding good input differences with higher accuracy. From the perspective of execution time, it showed an improvement of approximately 30%.

For web security, it’s essential to accurately classify traffic across various web applications to detect malicious activities lurking within network traffic. However, the encryption protocols for privacy protection, such as TLS 1.3 and IPSec, make it difficult to apply traditional traffic classification methods like deep packet inspection (DPI). Recently, the advent of deep learning has significantly advanced the field of encrypted traffic analysis (ETA), outperforming traditional traffic analysis approaches. Notably, pre-trained deep learning based ETA models have demonstrated superior analytical capabilities. However, the security aspects of these deep learning models are often overlooked during the design and development process. In this paper, we conducted adversarial attacks to evaluate the security of pre-trained ETA models. We targeted ET-BERT, a state-of-the-art model demonstrating superior performance, to generate adversarial traffic examples. To carry out the adversarial example generation, we drew inspiration from adversarial attacks on discrete data, such as natural language, defining fluency from a network traffic perspective and proposing a new attack algorithm that can preserve this fluency. Finally, in our experiments, we showed our target model is vulnerable to the proposed adversarial attacks.