In the increasingly complex domain of Korean voice phishing attacks, advanced detection techniques are paramount. Traditional methods have achieved some degree of success. However, they often fail to detect sophisticated voice phishing attacks, highlighting an urgent need for enhanced approaches to improve detection performance. Addressing this, we have designed and implemented a novel artificial neural network (ANN) architecture that successfully combines data-centric and model-centric AI methodologies for detecting Korean voice phishing attacks. This paper presents our unique hybrid architecture, consisting of a 1-dimensional Convolutional Neural Network (1D CNN), a Bidirectional Long Short-Term Memory (BiLSTM), and Hierarchical Attention Networks (HANs). Our evaluations using the real-world KorCCVi v2 dataset demonstrate that the proposed architecture effectively leverages the strengths of CNN and BiLSTM to extract and learn contextually rich features from word embedding vectors. Additionally, implementing word and sentence attention mechanisms from HANs enhances the model’s focus on crucial features, considerably improving detection performance. Achieving an accuracy score of 99.32% and an F1 score of 99.31%, our model surpasses all baseline models we trained, outperforms several existing solutions, and maintains comparable performance to others. The findings of this study underscore the potential of hybrid neural network architectures in improving voice phishing detection in the Korean language and pave the way for future research. This could involve refining and expanding upon this model to tackle increasingly sophisticated voice phishing strategies effectively or utilizing larger datasets.

With the widespread global trend of voice phishing or vishing attacks, the development of effective detection models using artificial intelligence (AI) has been hindered by the lack of high-quality and large volumes of data. This lack of data reflecting a real vishing scenario often leads to imbalanced datasets and biased detection models. Therefore, we present in this paper a data augmentation (DA) method for expanding the imbalanced Korean call content vishing (KorCCVi) dataset to address the existing data asymmetry problem and enhance the performance of Korean vishing detection. The proposed approach for DA involves using the back-translation (BT) method with three different intermediate languages: English, Chinese, and Japanese. The proposed method offers several advantages over the traditional synthetic minority oversampling technique (SMOTE), which is the main technique used to compare with our multilingual BT approach. Using these two DA techniques, several machine learning (ML) and deep learning (DL) models were trained on the original imbalanced dataset, the dataset balanced with SMOTE and its variants, and the dataset augmented with our method. We analyzed the impact of these DA methods on the performance of the models, demonstrated the benefits of each approach, and suggested the most suitable approach. The performance of the trained models was evaluated using the accuracy, precision, recall, and F1-score metrics. The experimental results demonstrated that the proposed multilingual BT method effectively expands the dataset while preserving its contextual and linguistic characteristics. The average performance of the models revealed that those trained on the augmented dataset outperformed the other models. They achieved F1-scores of 98.91% for the back-translated data, 98.14% for the original data, and 97.23% for SMOTE.