Enhancing cyber-attack prediction through optimized feature representation and advanced learning techniques

The integrity of computer networks and user security faces severe threats from web application attacks. Current threat detection techniques primarily rely on signature-based approaches, limiting their ability to recognize zero-day vulnerabilities. Moreover, the lack of comprehensive statistics on actual cyber-attacks further diminishes the effectiveness of these strategies. This paper introduces a comprehensive four-step methodology along with an architectural framework for the development of a robust cyberattack threat intelligence strategy. The initial phase involves data acquisition, encompassing the gathering of network traffic information and web page crawls, enabling the creation of feature vectors that effectively characterize cyber-attack information. Subsequently, the utilization of a sparse auto-encoder facilitates the analysis of the identified attack features. Finally, the proposed methodology incorporates the Convolutional Neural Network (ConvNNet) technique for systematic attack class prediction. Anomaly detection techniques are applied to forecast web-based attacks. The assessment leverages online cyber-attack datasets to evaluate the effectiveness of the proposed model. The original data yields a detection rate (DR) of 98.5% and a False Alarm Rate (FAR) of 9.5%. With training data, the model demonstrates an improved DR of 99% and a reduced FAR of 2%. Empirical analyses highlight the superior performance of the suggested approach compared to four competing machine learning methods, as evidenced by detection and false alarm rates across real-world and simulated web data