Algorithmic Approaches and its Analysis for the Detection of Deepfakes in Audio
Akruti Sarangi, C S Veeksha, K. Mounika, H Tejaswini, Uma R, Deepthi Shetty · 2025
Voice deepfakes pose a critical challenge in audio authentication, deceiving human listeners and automated systems. This study compares algorithmic approaches to detect voice deepfakes, aiming to enhance authentication system robustness. Evaluated models include LSTM, CNN, GBM, SVM, LCB net, Wirenet, Logistic Regression, KNN, AdaBoost, XGBoost, and CatBoost. Each algorithm’s performance in test accuracy and, where applicable, test loss across specified training epochs is assessed. Results indicate XGBoost achieves the highest accuracy at 99.3208828%, followed closely by GBM at 98.81%, CNN at 98.50%, and LSTM at 97.71%. In contrast, Logistic Regression and KNN exhibit lower accuracies of 74.92% and 80.65%, respectively. These findings underscore the effectiveness of gradient boosting and deep learning in detecting subtle voice manipulations, which are essential for advancing voice authentication security measures.