Detection methodology & specifications
Technical documentation on our forensic neural pipeline, spatial-frequency analysis, and operational limits.
1. Forensic philosophy
A categorical “100% Real” or absolute authenticity verdict overstates what digital signal forensics can mathematically prove. Clean media indicates an absence of synthetic diffusion signatures, face-swap boundary seams, or neural vocoder phase artifacts.
Sealed Rose delivers calibrated probabilistic outcomes divided into three principal findings:
Sensor noise, physical optics, and EXIF metadata align with natural hardware capture.
Severe compression blur or low resolution prevents extracting clean forensic indicators.
Statistical signatures matching diffusion noise or neural vocoder patterns detected.
2. Multi-stage pipeline
Our inspection architecture runs through three decoupled stages:
- Ingestion & Container Demuxing: Media is inspected in memory. Streams are demuxed into uncompressed audio waveforms and raw video frames. Header metadata is parsed for software encoding signatures.
- Ensemble Neural Classification: Frames and acoustic spectrograms are evaluated across specialized classification models, including primary neural detection models, spatial artifact classifiers, and custom multi-model ensembles.
- Confidence Calibration: Individual model logits are combined using empirical weighting to output a unified AI likelihood score between 0% and 100%.
3. Spatial and temporal checks
For video files, individual frame checks alone are vulnerable to false positives caused by heavy motion blur. Sealed Rose utilizes temporal consistency sampling across the media duration:
- Inter-frame Coherence: Diffusion models frequently exhibit micro-jitter in fine geometry (fingers, pupils, text, and hairline boundaries) from one frame to the next.
- Audio-Visual Synchronization: Phoneme-to-viseme lip alignment is verified against speech spectrogram formants.
- Fourier Spectral Analysis: High-frequency Fourier decomposition reveals characteristic periodic checkerboard artifacts left behind by generative upsampling layers.
4. Latency and processing benchmarks
Our serverless inference nodes are provisioned with GPU acceleration for rapid response times:
| Endpoint | Processing latency | Credit cost |
|---|---|---|
| Image detection | 500–800 ms | 1 credit |
| Audio detection | 0.6–1.2 s | 1 credit |
| Video detection | 2.5–4.5 s | 5 credits |
| Text detection | Not yet measured | 1 credit |
| Phone lookup | < 250 ms | 1 credit |
5. Compression and operational limitations
No synthetic media detection system is completely immune to degradation. In particular, repeatedly recompressed media (e.g. videos sent through WhatsApp, TikTok, or Twitter multiple times) loses fine-grained high-frequency pixel noise. When sensor noise is thoroughly destroyed, our pipeline categorizes the outcome as Inconclusive rather than issuing an ungrounded verdict.
For further inquiries regarding enterprise throughput, model evaluation, or forensic research, contact our team at supportsealedrose.com.