TDCGAN: A Novel GAN-Based Framework for Addressing Class Imbalance in Phishing Attack Detection

Phishing attacks remain a significant cybersecurity challenge, necessitating innovative detection methods to distinguish legitimate URLs from malicious ones. Despite numerous proposed solutions, many still encounter limitations. This study introduces a novel framework that integrates traditional machine learning techniques with a Generative Adversarial Network (GAN) for URL classification, focusing specifically on addressing class imbalance challenges. Our approach includes a rigorous feature selection process, where Random Forest outperformed Gradient Boosting and Decision Trees in identifying the top 10 most influential features. Additionally, we introduce the Triple Discriminator Conditional GAN (TDCGAN) as an innovative method to balance data. Experimental findings demonstrate that TDCGAN significantly enhances classification performance, achieving superior met-rics-87.21 % accuracy, 87.75% F1 Score, 84.01 % precision, and 91.85% recall-compared to traditional methods like SMOTE and RandomOverSampler. In the classification task, our proposed solution surpassed benchmarks such as K-Nearest Neighbors (KNN) and Random-ForestClassifier, achieving 93.76% accuracy, 88.42% F1 Score, and 93.64 % recall. These outcomes underscore the efficacy of our GAN-based approach in enhancing the accuracy and robustness of phishing detection systems.