Performance Analysis of Machine Learning and Deep Learning Algorithms for Detecting Web Attacks

Abstract

In today’s hyperconnected digital ecosystem, web applications act like bustling marketplaces, constantly exchanging data, services, and user interactions. However, this openness also invites a swarm of cyber threats such as SQL injection, cross-site scripting (XSS), denial-of-service (DoS), and other malicious activities. Traditional rule-based security systems struggle to adapt to evolving attack patterns, necessitating intelligent and adaptive detection mechanisms. This research presents a comprehensive performance analysis of various Machine Learning (ML) and Deep Learning (DL) algorithms for detecting web-based attacks using structured datasets.The proposed system integrates multiple algorithms including Support Vector Machine (SVM), Naïve Bayes, Autoencoder, and Long Short-Term Memory (LSTM) networks to identify abnormal patterns in web traffic. Initially, the dataset is preprocessed using encoding techniques such as Label Encoding and One-Hot Encoding to transform categorical features into numerical representations suitable for model training. The dataset is then split into training and testing subsets to evaluate model performance effectively.The study begins with traditional ML algorithms like SVM and Naïve Bayes, which are trained on labeled data to classify traffic as normal or abnormal. These models serve as baseline classifiers due to their simplicity and efficiency. The proposed system further enhances detection capabilities by incorporating an Autoencoder, a deep learning model used for anomaly detection through reconstruction error analysis. This unsupervised approach identifies deviations from normal behavior, making it effective for detecting previously unseen attacks.Additionally, an LSTM-based model is implemented as an extension to capture temporal dependencies in sequential data, which is crucial for understanding patterns in network traffic over time. Performance metrics such as accuracy, precision, recall, and F1-score are used to evaluate and compare the effectiveness of each algorithm.Experimental results demonstrate that deep learning models, particularly Autoencoders and LSTM, outperform traditional ML approaches in detecting complex and unknown attack patterns. The system also provides a user-friendly web interface built using Django, allowing users to upload datasets, train models, visualize performance metrics, and predict attack types in real-time.Overall, this research highlights the importance of combining multiple ML and DL techniques to build a robust and scalable web attack detection system. The findings suggest that hybrid approaches can significantly improve detection accuracy and adaptability, making them suitable for modern cybersecurity applications.