AI Music DeepFake Detection using Hybrid Autoencoder-Transformer Architecture

This paper presents a hybrid deep learning approach combining autoencoders with transformer architectures to detect AI-generated music. The system achieves 95% accuracy on a balanced dataset of 400 samples (200 authentic GTZAN tracks and 200 Suno AI-generated tracks), with 100% recall on real music (zero false negatives) and 90% recall on AI-generated content. The model consists of 21.1 million parameters and employs a combined loss function (70% classification + 30% reconstruction) to learn both discriminative and reconstructive features from mel-spectrogram representations. Trained on NVIDIA GeForce MX450 with PyTorch 2.7.1, the model reaches optimal performance at epoch 27 with early stopping.