A Quantum-Hybrid Framework for Enhanced Deepfake Detection: Integrating QAOA-Based Feature Selection With Quantum-Inspired Attention Mechanisms

Deepfake technology has created major new challenges to the authentication of digital media and the security of digital information. The traditional methods using deep learning to detect deepfakes have shown good results. However, many of these traditional methods suffer from redundant features, imbalanced classes, and poor interpretability. This paper introduces a new hybrid quantum-classical framework for detecting deepfakes that combines InceptionResnetV1 to extract high dimensional representation with a Convolutional Neural Network (CNN), and a Quantum Inspired Attention Mechanism to select the most important features to process. Additionally, this paper proposes a Quantum Approximate Optimization Algorithm (QAOA) for feature selection based on high dimensional representations of images obtained from an InceptionResnetV1. QAOA selects the most informative features to be input into the CNN and helps reduce dimensionality while maintaining the ability to discriminate between classes of data. The selected features are then passed to a Multi-Head Quantum Inspired Attention Layer that utilizes quantum interference patterns to re-weight feature importance dynamically. The use of a focal Loss function with label smoothing and confidence penalty is also proposed to balance the class distribution in the deepfake datasets. Experiments conducted on the FaceForensics++ benchmark dataset show that the quantum-hybrid method outperforms classical methods in terms of accuracy, precision, recall, and calibration metrics. Grad-CAM is used as a visualization tool to provide interpretable output by showing which facial areas the model is focusing on when identifying forensic evidence.