QCNN: Quantum Convolutional Neural Networks for Audio Deepfake Detection

The rapid diffusion of audio deepfakes poses a significant threat to the trustworthiness of multimedia content, social media platforms, and digital communication systems. Recent advances in generative models have enabled the creation of highly realistic synthetic speech, making reliable detection increasingly challenging for classical deep learning approaches. This study explores the application of Quantum Convolutional Neural Networks (QCNNs) for audio deepfake detection by exploiting quantum feature encoding and hierarchical quantum processing. Raw audio signals are transformed into Mel spectrograms and subsequently compressed into compact feature representations suitable for quantum encoding. An 8-qubit ZFeatureMap is used to map classical spectral features into a quantum state, which is then processed through a QCNN architecture composed of quantum convolution and pooling layers. Model parameters are optimized using the gradient-free Constrained Optimization BY Linear Approximations (COBYLA) optimizer within a hybrid quantum-classical learning framework. The proposed approach is evaluated on the FakeAVCelebV2 dataset, focusing exclusively on the audio modality. Experimental results indicate that the QCNN achieves a training accuracy of 81.06% and a test accuracy of 84%, with balanced performance across real and fake audio samples. An analysis of quantum feature distributions further suggests that the model effectively captures discriminative spectral patterns relevant to deepfake detection. Overall, these findings highlight the potential of QCNNs as a resource-efficient and reliable solution for audio forensics and contribute to ongoing research on quantum machine learning for trustworthy multimedia analysis.