A lightweight feature extraction technique for deepfake audio detection

The emergence of audio deepfakes has prompted concerns over reputational integrity and dependability. Deepfakes with audio can now be produced more easily, which makes it harder to spot them. Technologies that can identify audio-level deepfakes must be developed in order to address this issue. As a result, we have recognised the importance of feature extraction for these systems and we have created an improved method for feature extraction. On audio Mel spectrogram, we have employed a modified ResNet50 to extract features. Then, Linear Discriminant Analysis (LDA) dimensionality reduction technique have been used to optimise the feature complexity. The chosen features by LDA are then utilised to train these machine learning (ML) models using the backend classification algorithms Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbour (KNN), and Naive Bayes (NB). The ASVspoof 2019 Logical Access (LA) partition is utilised for training, ASVspoof 2021 deep fake partition are used to evaluate the systems. Also, we have used DECRO dataset for evakuating our proposed model under unseen noisy dataset. We have used 20% audios from training dataset for validation purpose. When compared to other models, our proposed method performs better than traditional feature extraction methods such as Mel Frequency Cepstral Coefficients (MFCC) and Gammatone Cepstral Coefficients (GTCC). It achieves an impressive Equal Error Rate (EER) of only 0.4% and an accuracy of 99.7%.