Neural Networks for Cryptographic Key Generation and Network-Traffic Analysis

Abstract

we studied the practical application of neural networks to cryptographic key generation and network traffic analysis. The latter task plays a key role in intrusion detection systems. We implemented our approach in Python using specialized libraries. For key generation, we employed a three-layer multilayer perceptron trained on random vectors containing ten random numbers each. The network produced a 128-bit binary sequence suitable for use as a cryptographic key. The entropy of the generated keys was estimated using Shannon’s formula; the resulting entropy was close to one, indicating the feasibility of the proposed approach. For network traffic analysis, we trained a model on a preprocessed dataset. The preprocessing reduced dimensionality and improved attack-detection accuracy. Each model was trained for one hundred epochs. Model performance was evaluated using the accuracy metric. The results showed no signs of overfitting, and the model achieved an accuracy exceeding 99%, demonstrating its suitability for network connection classification.