Robust, Explainable, and Multimodal Deepfake Detection Using Machine Learning for Digital Security: A Review

Abstract

Deepfake-enabled fraud caused an estimated $200 million in financial losses in the first four months of 2025 alone, and the same generation technology is increasingly aimed not only at public figures but at everyday individuals, a shift that has already prompted new federal legislation in the United States. Detecting synthetic media reliably enough to matter for digital security requires progress on three fronts at once: robustness to manipulation techniques and datasets a model was never trained on, explainability sufficient for a human analyst or, increasingly, a court to act on a flagged result, and the ability to fuse evidence across visual, audio, and other modalities rather than relying on any single one. This review synthesizes seven studies (2024–2025) organized around exactly these three requirements. On robustness, a parameter-efficient fine-tuning method reaches 91.2% average cross-dataset AUROC across fourteen benchmark datasets spanning 2019–2025, outperforming the next-best method by 2.7 points and showing that generalization, not raw benchmark accuracy, is the harder and more consequential problem. On multimodality, audio-visual fusion methods reach 95.24% and 95.5% accuracy/AUC respectively by exploiting inconsistencies between a face's visual emotion and a voice's acoustic and semantic content, and a broader survey documents the field's shift from single-modality detection toward the fusion architectures these applied studies exemplify. On explainability, a systematic review categorizes the field's growing toolkit — from forensic artifact analysis to Grad-CAM and SHAP to natural-language explanations — while noting that few studies test explanations against the same adversarial scrutiny applied to the detectors themselves. Two further studies, still emerging and not yet independently verifiable on their quantitative results, illustrate this field's reach beyond pure computer vision: a multi-classifier face-detection pipeline, and a framing of deepfake detection as an organizational digital-security and business-analytics concern for management information systems. Read together, these seven studies suggest that robustness, explainability, and multimodal fusion are converging demands on any deepfake detector meant for real deployment, not separable research tracks.