INTRUSION DETECTION OF NOVEL POLYMORPHIC CREDENTIAL SPRAYING ATTACKS BY MACHINE LEARNING APPLICATIONS IN CYBER SECURITY

The increasing sophistication of cyber threats requires robust intrusion detection systems capable of identifying known and unknown attacks. This study investigates the generalization capabilities of machine learning (ML) and deep learning (DL) models when applied to a newly simulated Polymorphic Credential Spraying (PCS) attack in a cloud environment. We trained multiple ML/DL classifiers on a balanced version of the CICIDS2017 dataset, including Random Forest, Gradient Boosting, LSTM, 1D CNN, and Autoencoder models. A novel PCS attack was simulated on an AWS-hosted SSH service, and the resulting network traffic was captured and preprocessed into feature vectors compatible with the trained models. Our evaluation highlights strengths and limitations in generalizing from known to novel attacks, offering insights into the feasibility of predictive security in dynamic, cloud-based infrastructures.