Voxear
Advanced deepfake detection combining metadata forensics, Xception spatial analysis, and physics-based verification for comprehensive video authenticity analysis
Detection Examples
Real sample videos from the test case folder, paired with backend-scored module outputs and verdict confidence.
Backend result: REAL. Metadata was uncertain, AI scored 27.0% fake, and physics found no anomalies.
#real-caseBackend result: FAKE. Metadata was uncertain, while AI and corneal reflection both flagged manipulation.
#deepfakeBackend result: FAKE. C2PA/content credentials indicated AI-generated content.
#metadataBackend result: FAKE. Metadata was uncertain, while AI and physics both produced fake signals.
#splicedMulti-Layered Detection Pipeline
Voxear implements a hierarchical detection pipeline combining three complementary approaches:
1. Metadata Forensics
First-line defense checking for generator watermarks and encoding anomalies.
2. Xception Spatial
Xception CNN analyzing face crops with per-frame scoring, temporal anomaly detection, and Grad-CAM spatial explainability.
3. Physics Verification
Independent checks for violations of physical laws in motion, lighting, and facial dynamics.
The system uses adaptive frame sampling and SHAP explainability to deliver transparent results suitable for forensic and research applications.
Detection Methods
Click any card to explore the technical details, algorithms, and visualizations behind each detection module.
AI Detection
Xception CNN trained on deepfake datasets, analyzing face crops with per-frame scoring, temporal anomaly detection, and Grad-CAM spatial explainability.
Click to learn more →
Physics-Based Detection
Analyzes physical properties and motion patterns to detect violations of natural laws that deepfakes often fail to replicate correctly.
Click to learn more →
Metadata Analysis
Examines video file metadata, encoding parameters, and digital signatures to detect AI-generated content markers and manipulation traces.
Click to learn more →
Explainability Features
Provides transparent, interpretable insights into detection results using SHAP values, attention weights, and visual explanations.
Click to learn more →