Video Deepfake Detection

Detect AI-generated and manipulated videos including face swaps, video replay attacks, and liveness spoofing to prevent video-based fraud.

Common Attack Vectors

Face Swap Videos

Real-Time Face Replacement:

• Live webcam feed with face replaced in real-time
• Software like DeepFaceLab, FaceSwap
• Can fool basic liveness checks
• Increasingly common in video KYC fraud

Pre-Recorded Face Swap:

• Legitimate user's verification video with face swapped
• Target person's face mapped onto real person's movements
• Synced with audio
• Very convincing if done well

Video Replay Attacks

Screen Playback:

• Playing pre-recorded video on phone/tablet screen
• Held up to webcam
• Can include real person performing liveness actions
• Detectable via screen moiré patterns

Virtual Camera:

• Software (OBS, ManyCam) feeds pre-recorded video to webcam
• No physical screen (harder to detect)
• Often used with deepfake face swap
• Can include interactive elements

AI-Generated Videos

Synthetic Person:

• Completely AI-generated person (doesn't exist)
• Generated frame-by-frame or video-to-video
• May include synthesized speech
• Very sophisticated attacks

Video-to-Video Translation:

• Target person's movements mapped to source video
• Maintains expressions and head movements
• Convincing for short clips
• Emerging threat

5 Credits Per video analysis

Detection Capabilities

1. Temporal Coherence Analysis

Analyze frame-to-frame consistency.

What We Check:

• Blinking Patterns: Natural vs. synthetic blinks
• Micro-Expressions: Subtle facial movements
• Head Movement: Natural motion patterns
• Background Consistency: Static background should remain static
• Lighting Changes: Abrupt vs. gradual changes

Deepfake Indicators:

• Unnatural blinking frequency (GANs struggle with eyes)
• Jittery movements (frame inconsistency)
• Face "pops" or sudden changes
• Background shifts when face moves (poor masking)

Analysis:

{
  "temporalCoherence": {
    "score": 45, // 0-100, lower = more suspicious
    "issues": [
      "Irregular blinking pattern (7 blinks in 3 seconds)",
      "Micro-jitter in face region",
      "Background shifts inconsistently with head movement"
    ],
    "suspiciousFrames": [45, 67, 89, 112]
  }
}

2. Face Detection Consistency

Track face detection confidence across frames.

Authentic Video: Face detected with 95-99% confidence in all frames

Deepfake Video: Confidence drops or face lost in some frames

• GAN artifacts confuse face detector
• Poorly blended face swap
• Temporal inconsistency

{
  "faceDetection": {
    "averageConfidence": 72, // Should be 95+
    "stableDetection": false,
    "framesWithLowConfidence": 23, // out of 150 frames
    "verdict": "suspicious"
  }
}

3. Optical Flow Analysis

Analyze motion patterns between frames.

Natural Motion:

• Smooth, continuous movement
• Consistent velocity
• Proper motion blur

Deepfake Motion:

• Abrupt changes in motion direction
• Inconsistent motion vectors
• Missing or incorrect motion blur

4. Audio-Visual Synchronization

Check if lip movements match audio.

Lip-Sync Analysis:

{
  "audioVisualSync": {
    "lipSyncScore": 68, // 0-100, authentic usually 90+
    "issues": [
      "Lip movements lag audio by 200ms",
      "Mouth shape doesn't match phonemes",
      "Audio quality inconsistent with video"
    ],
    "verdict": "likely_deepfake"
  }
}

Common in:

• Voice cloning + face swap attacks
• Video with dubbed audio
• Screen replay with separate audio track

5. Screen Replay Detection

Detect videos played on screens and re-captured.

Detection Methods:

• Moiré Patterns: Interference from screen pixels
• Bezels/Edges: Screen edges visible in frame
• Refresh Rate Artifacts: Screen flicker at specific frequencies
• Reflection Analysis: Screen reflections
• Resolution Mismatch: Lower resolution from screen capture

{
  "screenReplay": {
    "detected": true,
    "confidence": 89,
    "indicators": [
      "Moiré pattern detected in frames 23-145",
      "Screen bezel visible in corner",
      "60Hz screen refresh detected",
      "Consistent rectangular region (device screen)"
    ]
  }
}

Liveness Detection

Passive Liveness

Analyze single frame or short video for spoofing.

Detection Methods:

• 3D depth estimation (flat photos lack depth)
• Texture analysis (paper/screen texture)
• Face movement naturalness
• Eye tracking (reflections, pupil dilation)

Result:

{
  "passiveLiveness": {
    "result": "fail",
    "confidence": 92,
    "spoofingType": "photo_attack",
    "indicators": [
      "No depth detected (flat surface)",
      "Paper texture visible",
      "No pupil dilation observed"
    ]
  }
}

Active Liveness

User performs specific actions (blink, smile, turn head).

Challenges:

1. "Please blink twice"
2. "Smile"
3. "Turn your head to the left"
4. "Turn your head to the right"

Analysis:

{
  "activeLiveness": {
    "result": "pass",
    "challenges": [
      { "action": "blink", "response": "correct", "naturalness": 94 },
      { "action": "smile", "response": "correct", "naturalness": 88 },
      { "action": "turn_left", "response": "correct", "naturalness": 91 },
      { "action": "turn_right", "response": "correct", "naturalness": 90 }
    ],
    "overallScore": 91
  }
}

Deepfake Challenges:

• Actions performed but movements unnatural
• Timing delays (deepfake processing lag)
• Artifacts during complex movements

Analysis Process

Step 1: Upload Video

const formData = new FormData();
formData.append('video', videoFile);
formData.append('analysisType', 'verification'); // or 'general'
formData.append('includeLiveness', 'true');
 
const job = await fetch('/api/v4/deepfake/analyse', {
  method: 'POST',
  headers: {
    'Authorization': `Bearer ${apiKey}`
  },
  body: formData
});
 
// Returns job ID (processing takes 30-120s)
{
  "jobId": "job_xyz789",
  "status": "processing",
  "estimatedTime": "45 seconds"
}

Step 2: Processing

Analysis Pipeline (30-120 seconds):

1. Video preprocessing (extract frames, audio)
2. Frame-by-frame analysis (150-300 frames)
3. Face detection and tracking
4. Temporal coherence analysis
5. Optical flow calculation
6. Audio-visual sync check
7. Liveness detection (if applicable)
8. Screen replay detection
9. Final risk scoring

Step 3: Receive Results

GET /api/v4/deepfake/jobs/:jobId
 
{
  "jobId": "job_xyz789",
  "status": "completed",
  "video": {
    "filename": "verification_video.mp4",
    "duration": "12.5 seconds",
    "frameCount": 375,
    "resolution": "1280x720",
    "fps": 30
  },
  "result": {
    "isDeepfake": true,
    "confidence": 89,
    "manipulationType": "face_swap",
    "riskScore": 85,
    "recommendation": "reject"
  },
  "analysis": {
    "temporalCoherence": {
      "score": 42,
      "issues": ["Irregular blinking", "Micro-jitter"]
    },
    "faceDetection": {
      "averageConfidence": 76,
      "stableDetection": false
    },
    "audioVisualSync": {
      "lipSyncScore": 65,
      "verdict": "poor_sync"
    },
    "screenReplay": {
      "detected": false
    },
    "liveness": {
      "result": "fail",
      "spoofingType": "deepfake_video"
    }
  },
  "suspiciousFrames": [
    {
      "frameNumber": 45,
      "timestamp": "1.5s",
      "issue": "Face detection dropped",
      "confidence": 42
    },
    {
      "frameNumber": 89,
      "timestamp": "3.0s",
      "issue": "Blending artifact",
      "confidence": 68
    }
  ]
}

Confidence Scores

Video Requirements

Technical Requirements

Quality Checks

Reject If:

• Resolution < 640x480
• FPS < 15
• Face too small (< 200x200px)
• Face not visible in >20% of frames
• Extreme lighting (overexposed/underexposed)
• Heavy motion blur throughout

Integration with Verification

Enhanced Verification Workflow

{
  "workflowType": "enhanced",
  "livenessType": "active",
  "enableDeepfakeCheck": true, // Video deepfake scan
  "videoRequirements": {
    "minDuration": 5, // seconds
    "maxDuration": 15,
    "requireActions": ["blink", "smile", "turn_head"],
    "rejectDeepfakeThreshold": 70
  }
}

Flow:

1. User records video with liveness actions
2. Video uploaded
3. Liveness check (user performed actions?)
4. Deepfake scan (is video authentic?)
5. Face match (matches ID document?)
6. Decision: Approve/Reject/Manual Review

Best Practices

1. Require active liveness - Harder to deepfake with random actions
2. Set video duration 5-15 seconds - Enough frames to analyse, not too long
3. Check audio-visual sync - Voice cloning often has sync issues
4. Combine with image deepfake - Scan ID document too
5. Manual review uncertain cases - 70-89% confidence range
6. Limit retries - Fraudsters will retry until successful, limit to 3 attempts
7. Monitor suspicious patterns - Multiple failed liveness from same device

Limitations

Detection Challenges

Current Performance:

• Face swap detection: 92%
• Screen replay: 97%
• Liveness spoofing: 94%
• Overall: 92%

Difficult Cases:

• High-quality real-time deepfakes (expensive hardware required)
• Short videos (<5 seconds, fewer frames to analyse)
• Low resolution (artifacts harder to detect)
• Poor lighting (masks deepfake artifacts)

False Positives

Common Causes:

• Webcam quality (low-end cameras have artifacts)
• Compression (heavy video compression adds artifacts)
• Lighting changes (moving between shadows)
• Network lag (in live video calls, can cause jitter)

Mitigation:

• Use 70%+ threshold
• Manual review 70-89% range
• Allow retake with better lighting/camera

Attack Evolution

Current Threats (2024)

1. Real-Time Face Swap:

   • Software: DeepFaceLive, FaceSwap Live
   • Requires: Powerful GPU, training data
   • Effectiveness: Medium (detectable with temporal analysis)

2. Video-to-Video:

   • Techniques: First Order Motion Model, Face2Face
   • Quality: High for short clips
   • Effectiveness: High (harder to detect)

3. Screen Replay:

   • Simple but common
   • Effectiveness: Low (easily detected)

Emerging Threats

• Neural Rendering: Real-time photorealistic face synthesis
• Audio-Driven Face Reenactment: Lip-sync from audio only
• Full-Body Deepfakes: Beyond just face replacement

API Reference

Analyze Video

POST /api/v4/deepfake/analyse
 
// multipart/form-data
{
  "video": File,
  "analysisType": "verification" | "general",
  "includeLiveness": boolean,
  "returnSuspiciousFrames": boolean
}
 
// Returns
{
  "jobId": "job_xyz789",
  "status": "processing",
  "estimatedTime": "45 seconds"
}

Get Results

GET /api/v4/deepfake/jobs/:jobId
 
// Returns full analysis including frame-by-frame data

Extract Suspicious Frames

GET /api/v4/deepfake/jobs/:jobId/frames?frameNumbers=45,89,112
 
// Returns specific frames as images for manual review

Pricing

Next Steps