DeepGuard-X: An Explainable, Uncertainty-Aware Multi-Modal Fusion Architecture for Deepfake and Manipulated Media Triage

Manipulated media evidence is now a routine feature of cybercrime casework, yet existing detection systems typically examine a single modality, return uncalibrated verdicts, and offer no mechanism for signaling their own uncertainty — a serious deficiency for any tool whose output may shape an investigation or inform legal proceedings. This paper presents DeepGuard-X, an explainable, uncertainty-aware, multi-modal fusion architecture for deepfake and manipulated-media triage, developed and evaluated under a strict held-out protocol in which no component — classifier, fusion weight, or uncertainty model — is ever scored on a case it observed during fitting. Independently calibrated image, video, and audio classifiers are combined through reliability-weighted log-odds fusion; Monte Carlo Dropout quantifies epistemic and aleatoric uncertainty; and non-circular, validation-selected thresholds drive a three-tier triage decision, with every verdict additionally backed by SHAP-based explainability. The architecture is evaluated on a newly constructed, literature-calibrated synthetic benchmark, DG-X SMFD v1.0 (50,000 cases, released publicly on Zenodo), designed to support reproducible methodology research pending access to licensed real-world deepfake corpora. On a genuinely held-out test split, the fused verdict achieves AUC-ROC = 0.9899 (95% CI [0.9878, 0.9918]), stable at 0.9894 ± 0.0004 across five independent seeds and substantially exceeding any individual modality (image 0.968, video 0.975, audio 0.863). A correlated-artifact stress test, an honestly-reported fusion-rule ablation, a variance-preserving adversarial evasion test, and a direct empirical comparison of epistemic and aleatoric uncertainty are all presented with findings reported exactly as observed — including an inconclusive fusion-rule comparison and a highly correlated (r = 0.91) uncertainty decomposition — rather than adjusted to fit a cleaner narrative. At its selected operating point, the triage layer resolves 60.1% of held-out cases automatically at 0.40% error, against a 4.91% unconditional baseline on the same split. Every result in this paper is explicitly scoped to synthetic data, and every identified limitation — including the absence of real-world validation, demographic fairness modeling, and human forensic-examiner evaluation — is discussed without exception. DeepGuard-X and its accompanying dataset are intended as a rigorously validated, fully reproducible reference architecture for future work incorporating genuine forensic media.