Multitask Deep Learning for Joint Deepfake Speech Detection and Spoofing Attack Attribution

Abstract The rapid advancement of neural speech synthesis and voice conversion technologies has increased the realism of synthetic speech, creating significant challenges for biometric authentication and digital forensics. This study proposes a Multi-Task Deep Learning (MT-DL) framework that jointly addresses binary deepfake speech detection and supervised spoof attribution of known attack classes. The framework combines complementary spectro-temporal representations (MFCC, Log-Mel spectrograms, and delta-MFCC features) with a convolutional front-end and a hybrid Bi-LSTM/Transformer temporal encoder to feed task-specific detection and attribution heads. Evaluated on the ASVspoof 2019 Logical Access dataset using a stratified 60:40 split, the framework achieves a binary detection accuracy of 99.75% (0.48% EER) and a supervised spoof attribution accuracy of 99.57% (99.58% Macro-F1) across known attacks (A01–A06). Unseen attacks (A07–A19) and cross-dataset evaluations on the In-the-Wild dataset further demonstrate robust generalization, though performance sensitivity under severe additive acoustic noise remains a key area for future optimization.