Robust Machine Learning under Data Poisoning and Model Attacks

This study tackled the issue of label-flipping attacks, a specific type of data poisoning in machine learning where malicious actors deliberately modify the class labels of training datasets to deceive models. These attacks pose a significant challenge for detection, as the characteristics of compromised records align with those of legitimate data, thereby jeopardizing the accuracy and dependability of models used in network intrusion detection systems. To counter this problem, a robust machine learning framework based on Random Forest was developed to identify and categorize mislabeled records within network traffic. The dataset comprised standard traffic alongside various attack categories, including Man in the Middle, with each class containing 1,229 records to maintain balance. The framework extracted 63 features from the network and utilized ensemble techniques to identify discrepancies between the features and their respective labels. Python was used for development, taking advantage of the libraries designed for data manipulation, model training, and performance evaluation. The findings indicated that the Random Forest model achieved an impressive accuracy of 98.47%, demonstrating strong precision, recall, and F1-scores for effectively identifying label-flipping attacks. The study emphasized that ensemble methods, such as Random Forest could uphold the model integrity in adversarial settings. Network security frameworks are advised to incorporate these robust machine learning solutions to protect against data-poisoning threats.