A Comparative Analysis of Backdoor and Label-Flipping Attacks on Credit Card Fraud Detection Using Ensemble Learning Techniques

Advanced machine learning models have been applied across various sectors, including healthcare, finance, agriculture, and transportation, yielding promising results. The continuously evolving nature of cybersecurity threats necessitates the development of intelligent systems to counteract these attacks, particularly adversarial data poisoning attacks that threaten the safety of deployment. Notably, backdoor attacks and labelflipping attacks compromise the integrity of predictive models. Backdoor attacks manipulate models to misclassify malicious inputs containing specific triggers, while label-flipping attacks alter class labels in training data to degrade overall model accuracy. These attacks continue to receive significant attention in fraud detection. This study investigates backdoor poisoning and two types of label-flipping attacks (random and strategic) on a credit card fraud detection dataset. The dataset is sourced from the Kaggle repository, consisting of transactions made by credit cardholders in September 2013. The class imbalance was addressed using the Synthetic Minority Oversampling Technique (SMOTE). A comparative analysis was carried out on 5 % of the fraudulent training data and 15% of the non-fraudulent training data poisoned using random label-flipping attacks, strategic labelflipping attacks, and backdoor poisoning. Ensemble learning models such as random forest, AdaBoost, and XGBoost were evaluated under clean and poisoned conditions. The results show that random forest, when subjected to backdoor poisoning, demonstrated the highest robustness, achieving an Area under the Precision-Recall Curve (AUPRC) of 0.89 under attack conditions. In contrast, AdaBoost, when faced with strategic label-flipping, proved to be the most vulnerable, with its AUPRC dropping to 0.62. Strategic label-flipping exhibited the most significant impact on model performance, confirming that it is the poisoning attack with the most significant impact.