Machine Learning-based Approach for Enhancing Multi-step Attack Prediction

Attack prediction is an advanced and crucial technology that enables response proactively for defending against impending attacks before compromising the target. Currently, most prediction methods rely on historical data and statistical methods to predict future threats. However, with the development of machine learning techniques, deep learning-based approaches hold promise for implementing data-intensive regression tasks in a constantly evolving environment. This project evaluates the effectiveness of two models, namely the Hidden Markov Model (HMM) and the Long Short-Term Memory (LSTM) model, in predicting multi-step attacks. By using the well-known NSL-KDD dataset and a real-world penetration dataset captured by ourselves, our experimental results reveal that the proposed model achieve promising results in forecasting subsequent attack phases, particularly when confronted with time-series data, with an accuracy of 80%. Additionally, the LSTM model demonstrates greater accuracy in single-step prediction than the HMM.