Technical Documentation

Detection Methodology

How GhostScreen detects AI-generated video in real-time. Our 5-method ensemble approach, benchmark results, and technical specifications.

You don't need to understand how we detect deepfakes to use GhostScreen. But if you're the kind of person who wants to know what's under the hood, this is for you.

We built our detection system on peer-reviewed research and real-world testing. Here's exactly how it works.

How It Works

GhostScreen uses a weighted ensemble of 5 independent detection methods, each analyzing different aspects of video authenticity. All processing happens locally in your browser using established computer vision techniques.

timer
Analysis Speed
<2 seconds per frame, runs every 6 seconds
face
Facial Landmarks
68-point detection via face-api.js (TinyFaceDetector)
lock
Privacy
100% local processing. Zero data transmitted.

Scoring System

70-100Likely Real Person
35-69Suspicious - Review Recommended
0-34Likely AI/Deepfake

Composite score calculated from weighted average of all 5 detection methods. Confidence increases with more signals detected.

5 Detection Methods

Each method targets specific artifacts common in AI-generated video. Combined, they provide robust detection across multiple deepfake generation techniques.

face_retouching_natural

1. Boundary Artifacts

25% Weight

Detects unnatural blending at face edges where AI-generated faces merge with backgrounds or original video.

Signals Detected

  • - Face edge blending artifacts
  • - Gaussian blur at boundaries
  • - Color discontinuities at jawline/hairline
  • - Alpha channel errors

Why It Works

DeepFaceLive, FaceSwap, and similar tools often produce visible seams where the synthetic face meets the original frame.

visibility

2. Eye Texture Analysis

20% Weight

Analyzes iris texture complexity and eye characteristics that AI often fails to reproduce accurately.

Signals Detected

  • - Iris texture entropy (threshold: 3.5)
  • - Specular reflection presence
  • - Pupil darkness levels
  • - Sclera brightness consistency

Why It Works

Real human irises have complex, high-entropy texture patterns (>3.5). AI-generated eyes often appear smoother with less detail.

timeline

3. Temporal Consistency

20% Weight

Tracks movement patterns across frames to detect the unnatural smoothness or glitches typical of AI interpolation.

Signals Detected

  • - Frame-to-frame movement variance
  • - Landmark tracking glitches (>15px jumps)
  • - Repetitive pattern detection
  • - Optical flow anomalies

Why It Works

Natural human movement has variance between 0.02-0.15 pixels per frame. AI interpolation is often too smooth or contains sudden jumps.

remove_red_eye

4. Blink Pattern Analysis

15% Weight

Monitors natural blinking behavior using Eye Aspect Ratio (EAR) calculations to detect synthetic patterns.

Signals Detected

  • - Blink rate (natural: 10-30/minute)
  • - Blink duration (natural: 100-400ms)
  • - Timing variance patterns
  • - EAR threshold: <0.2 = closed

Why It Works

Many deepfakes trained on still images fail to reproduce natural blinking. AI avatars often blink too regularly or too rarely.

light_mode

5. Lighting Physics

20% Weight

Validates that face lighting is physically consistent with the environment and follows natural illumination patterns.

Signals Detected

  • - Face-to-background lighting ratio (0.5-2.0 natural)
  • - Vertical gradient (top-to-bottom: 0.8-1.5)
  • - Shadow asymmetry
  • - Color grading mismatches

Why It Works

When a synthetic face is composited onto real video, the lighting often doesn't match the environment.

Benchmark Methodology

How we test and validate detection accuracy across multiple deepfake generation methods.

Test Dataset Composition

Real Person Videos40%
Face Swap (DeepFaceLive, Roop)25%
AI Avatars (HeyGen, Synthesia, D-ID)25%
Lip Sync (SadTalker, Wav2Lip)10%

Test Conditions

check_circle
Variable Lighting
Office, home, outdoor, studio conditions
check_circle
Multiple Resolutions
480p, 720p, 1080p webcam quality
check_circle
Compression Artifacts
Video call compression simulated
check_circle
Movement Variation
Static, head movement, gesturing

Accuracy Results

Current performance metrics from internal testing. We're transparent about what our technology can and cannot do.

9
Detection Methods
Multi-method ensemble
96%
Face Swap Detection
DeepFaceLive, Roop, FaceSwap
93%
AI Avatar Detection
HeyGen, Synthesia, D-ID
<3%
False Positive Rate
Real people flagged incorrectly
info

Honest Limitations

  • - Detection accuracy decreases with low-quality video (<480p) or extreme lighting conditions
  • - Sophisticated, well-lit AI avatars may occasionally pass detection
  • - Natural face filters and virtual backgrounds can trigger false positives
  • - Performance depends on consistent face visibility during the call
  • - We recommend using detection results as one signal among many in your hiring decision

Technical Specifications

Processing

Analysis Interval6 seconds
Frame Processing<2 seconds
Input Resolution224x224px
Frame History25 frames

Models

Face DetectionTinyFaceDetector
Landmarks68-point model
Model Size (Face)~189KB
Model Size (Landmarks)~349KB

Privacy & Compliance

Data StorageNone
Cloud UploadNone
GDPR CompliantYes
CCPA CompliantYes

Ready to see it in action?

Schedule a demo to see GhostScreen detect AI-generated video in real-time.