Robustness-Aware QAOA Feature Selection for Adversarially Resilient Deepfake Detection

This paper proposes a robustness-aware Quantum Approximate Optimization Algorithm (QAOA) feature-selection approach for adversarially resilient deepfake detection. The proposed framework extends an existing quantum-hybrid deepfake detection pipeline by incorporating adversarial feature stability into the feature-selection objective. The base framework combines InceptionResNetV1 for facial feature extraction, QAOA-based feature selection, quantum-inspired attention, balanced focal loss, and Grad-CAM for interpretability. The proposed extension aims to select a compact set of facial features that balances discriminative relevance, feature redundancy, subset size, and stability under controlled perturbations. The methodology introduces a feature-stability score and incorporates it into the QAOA cost function. The selected features are subsequently processed by the existing quantum-inspired attention mechanism. The proposed experimental design compares the original QAOA feature-selection method with the robustness-aware approach using consistent data partitions, preprocessing, feature extraction, attention, and classification components. Evaluation is planned using accuracy, precision, recall, F1-score, AUC-ROC, performance degradation under perturbations, calibration, feature stability, and feature-subset compactness. This work presents a proposed research framework and experimental methodology. Implementation results and measured performance improvements have not yet been reported. Keywords: Deepfake Detection, Quantum Approximate Optimization Algorithm (QAOA), Adversarial Robustness, Feature Selection, Quantum-Inspired Attention, Deep Learning, Grad-CAM.