Preventing the Impact of AI Deepfakes on Journalistic Integrity: A Hybrid Approach to Detection and Prevention

Abstract

Deepfakes pose a direct challenge to journalistic verification because convincing synthetic video and audio can circulate faster than manual fact-checking, while many high-performing detectors remain costly or lose reliability under compression and domain shift. This study presents a newsroom-oriented hybrid framework that combines lightweight visual screening, temporal consistency analysis, selective frequency-domain verification, optional audio analysis, content-provenance checks, and human editorial review. MobileNetV3 provides low-cost frame triage, a shallow GRU/LSTM models short-term temporal inconsistencies, and the frequency-aware stage is reserved for ambiguous content to limit computational overhead. Available branch scores are combined through late fusion, and uncertain or high-risk items are escalated to editors. The reported model comparison yields 91.8% accuracy, 89.4% precision, 90.5% recall, and an 89.9% F1-score for the MobileNet-LSTM configuration, exceeding the three internal baselines reported in this study. The contribution is not a claim of superior benchmark accuracy over all state-of-the-art detectors; rather, it is an operational design that integrates efficient detection, provenance evidence, multimodal checks, and editorial judgment for resource-constrained newsroom workflows.