How Krites Works
A multi-signal forensic approach to deepfake detection with transparent, interpretable analysis. Unlike black-box solutions, we show you exactly how and why media was flagged.
The Detection Pipeline
1
Upload
Upload image, video, or audio file
2
Process
Extract frames and audio tracks
3
Analyze
Run three detection signals
4
Report
Generate forensic report
Detection Signals
Visual Analysis
EfficientNet-B7 • GoogleState-of-the-art image classification that detects facial artifacts, lighting inconsistencies, and unnatural textures.
Face boundary artifactsLighting inconsistenciesGAN texture patternsBlending anomalies
Audio Analysis
Wav2Vec 2.0 • MetaSelf-supervised speech model that identifies synthetic audio patterns and voice cloning artifacts.
Spectral anomaliesSynthetic breathingUnnatural prosodyMissing micro-variations
Lip-Sync Analysis
AV-HuBERT • MetaAudio-visual speech model that measures synchronization between lip movements and audio.
A/V sync offsetPhoneme mismatchTemporal inconsistencyMouth movement artifacts
Why Multi-Signal Detection?
Single-Model Limitations
- • Fails on unfamiliar deepfake methods
- • Easily fooled by post-processing
- • No explanation of detection
- • High false positive rates
Multi-Signal Advantages
- • Robust against new techniques
- • Cross-validation between signals
- • Detailed, interpretable explanations
- • Higher accuracy and confidence