Cybersecurity Attack Detection Using LSTM and ResNet-50 Hybrid Model with Cloud Deployment

Abstract

As the number of complex cybersecurity threats accelerates at a fast pace, real-time attack identification
is now essential in defending computer systems. These methods are challenged by high-dimensional data, class
imbalance, and low scalability, resulting in low accuracy and increased false positives. This paper introduces a
hybrid DL archetype combining ResNet-50 for feature extraction and LSTM in learning lay sequences, running
on AWS Sage Maker for scalable cloud inference and training. The method involves extensive data pre-
processing, including imputation, one-hot encoding, and feature scaling, with model training optimized through
the Adam algorithm. Experimental results show excellent performance, with 99.50% of accuracy, 99.20% of
precision, 99.30% of recall, 99.40% of F1-score, and an AUC of 0.999998. These findings demonstrate the
strength and efficiency of the framework in real-time detection of cybersecurity attacks, providing an
enhancement over previous solutions through the ability to efficiently process intricate data patterns and deploy
quickly in cloud environments.