Image Deepfake Detection

Detect AI-generated images, photoshopped documents, and manipulated photos to prevent identity fraud and document forgery.

Common Attack Vectors

Fake ID Documents

Face Swap:

• Real ID document with different person's face pasted on
• Most common fraud type in identity verification
• Can be very convincing if done well

Template Forgery:

• Completely fabricated ID using templates
• AI-generated document with synthetic data
• May include AI-generated face photo

Photo Substitution:

• Original document with photo replaced
• Physical photo swap scanned
• Digital manipulation of photo area

AI-Generated Selfies

Synthetic Faces:

• Completely AI-generated person (doesn't exist)
• Generated by GANs (Generative Adversarial Networks)
• Often used to bypass face match checks

Face Swaps:

• Real person's face replaced with target's face
• Used to impersonate document holder
• Can fool basic face matching

Manipulated Documents

Data Changes:

• Altered dates of birth
• Changed names
• Modified addresses
• Fake document numbers

Photoshop Manipulation:

• Clone stamp to duplicate elements
• Content-aware fill to remove details
• Layer-based editing

2 Credits Per image analysis

Detection Capabilities

1. GAN Fingerprinting

Detect AI-generated images by identifying model-specific patterns.

How It Works:

• Different AI models (StyleGAN, ProGAN, DALL-E) leave unique "fingerprints"
• Analyze frequency domain for generation artifacts
• Detect unnatural color distributions
• Identify spectral patterns inconsistent with cameras

Accuracy: 96% for known GAN models

Detection Example:

{
  "isDeepfake": true,
  "manipulationType": "gan_generated",
  "model": "StyleGAN2",
  "confidence": 94,
  "artifacts": [
    "Spectral anomaly in high-frequency domain",
    "Unnatural color distribution",
    "Grid pattern artifacts (common in GANs)"
  ]
}

2. Photoshop Detection

Identify images edited with photo manipulation software.

Detection Methods:

• Metadata Analysis: EXIF data inconsistencies
• Compression Artifacts: Double JPEG compression
• Clone Detection: Duplicated image regions
• Noise Inconsistency: Different noise patterns in regions
• Lighting Analysis: Inconsistent light sources

Example:

{
  "isDeepfake": true,
  "manipulationType": "photoshop_edit",
  "confidence": 91,
  "artifacts": [
    "Clone stamp detected in photo area",
    "Double JPEG compression around face",
    "Inconsistent noise patterns",
    "Lighting mismatch between face and background"
  ],
  "editedRegions": [
    {
      "area": "photo_section",
      "bbox": { "x": 120, "y": 80, "width": 180, "height": 220 },
      "manipulation": "photo_replacement"
    }
  ]
}

3. Face Swap Detection

Detect when a face has been digitally replaced.

Detection Techniques:

• Blending Artifacts: Unnatural edges around face
• Lighting Inconsistency: Face lit differently than rest of image
• Resolution Mismatch: Face different resolution than background
• Skin Texture: Unnatural or inconsistent skin texture
• Face Landmarks: Distorted facial geometry

Common in:

• Fake ID documents
• Profile picture fraud
• Video call impersonation attempts

4. Screen Capture Detection

Detect when image is screenshot of another screen (not original photo).

Detection Signals:

• Moiré patterns (interference patterns from screen pixels)
• Screen bezels or UI elements visible
• Pixel grid patterns
• Lower resolution than expected
• Compression artifacts from screen capture

Why This Matters:

• Indicates user photographed another screen instead of real document
• Often means they don't possess physical document
• Common fraud technique

Image Quality Assessment

Before deepfake analysis, check image quality:

{
  "quality": {
    "resolution": { "width": 1920, "height": 1080, "score": 95 },
    "sharpness": 88, // Blur detection
    "lighting": 82, // Even lighting
    "contrast": 90,
    "colorBalance": 85,
    "overallScore": 86
  },
  "issues": [
    "Slight motion blur detected",
    "Uneven lighting (shadow on left side)"
  ]
}

Reject if:

• Resolution < 640x480
• Sharpness < 50 (too blurry)
• Overexposed (brightness > 240)
• Underexposed (brightness < 15)

Analysis Process

Step 1: Upload Image

const formData = new FormData();
formData.append('image', imageFile);
formData.append('analysisType', 'document'); // or 'selfie', 'generic'
 
const job = await fetch('/api/v4/deepfake/analyse', {
  method: 'POST',
  headers: {
    'Authorization': `Bearer ${apiKey}`
  },
  body: formData
});
 
// Returns job ID
{
  "jobId": "job_abc123",
  "status": "processing"
}

Step 2: Processing

Analysis Pipeline (2-5 seconds):

1. Image preprocessing (resize, normalise)
2. Quality assessment
3. GAN fingerprint analysis
4. Photoshop manipulation detection
5. Face detection and analysis (if applicable)
6. Metadata extraction and validation
7. Final risk scoring

Step 3: Receive Results

GET /api/v4/deepfake/jobs/:jobId
 
{
  "jobId": "job_abc123",
  "status": "completed",
  "image": {
    "type": "document",
    "filename": "id_card.jpg",
    "size": "2.3 MB",
    "resolution": "1920x1080"
  },
  "result": {
    "isDeepfake": true,
    "confidence": 94,
    "manipulationType": "face_swap",
    "riskScore": 88,
    "recommendation": "reject"
  },
  "analysis": {
    "ganFingerprint": {
      "detected": false,
      "confidence": 12
    },
    "photoshopEdit": {
      "detected": true,
      "confidence": 92,
      "method": "face_replacement"
    },
    "faceSwap": {
      "detected": true,
      "confidence": 94,
      "artifacts": [
        "Blending artifacts around face boundary",
        "Lighting inconsistency",
        "Resolution mismatch"
      ]
    },
    "screenCapture": {
      "detected": false
    }
  },
  "metadata": {
    "camera": "Unknown (stripped)",
    "dateTime": "2024:01:10 14:32:10",
    "software": "Adobe Photoshop 2023",
    "gps": null
  }
}

Confidence Scores

Common Deepfake Indicators

For ID Documents

• Clone stamp patterns: Duplicated texture in background
• Mismatched lighting: Face lit differently than document
• Inconsistent security features: Holograms that don't match template
• Font irregularities: Wrong font or size for document type
• Alignment issues: Photo not properly aligned in frame
• Edge artifacts: Unnatural edges around replaced photo

For Selfies

• Synthetic skin texture: Too smooth or artificially detailed
• Unnatural eyes: Asymmetric pupils, odd reflections
• Inconsistent hair: Hair blends unnaturally with background
• Facial distortions: Slight warping near face edges
• Background anomalies: Blurred or synthetic-looking background
• Color banding: Unnatural color gradients

Integration with Identity Verification

Automatic Deepfake Scanning

// Enable automatic deepfake check in verification workflow
{
  "workflowType": "enhanced",
  "enableDeepfakeCheck": true, // Adds deepfake scan
  "deepfakeThreshold": 70 // Reject if confidence >= 70
}

Workflow:

1. User uploads ID document
2. Document verified for authenticity
3. Deepfake scan performed ← Automatic
4. If deepfake detected → Reject or manual review
5. If clean → Continue with face match

Manual Deepfake Scan

Trigger deepfake scan on existing images:

POST /api/v4/deepfake/scan-verification
 
{
  "verificationId": "ver_xyz789",
  "scanDocument": true,
  "scanSelfie": true
}

Limitations

Detection Accuracy

Current Performance:

• Known GAN models: 96%
• Photoshop edits: 93%
• Face swaps: 92%
• Overall: 94%

Challenges:

• Very high-quality deepfakes (professional editing)
• New/unknown AI models
• Subtle manipulations (small data changes)
• Low-resolution images (harder to detect artifacts)

False Positives

Common Causes:

• Heavy makeup or filters (Instagram/Snapchat)
• Professional photo retouching (legitimate)
• Lens distortion (wide-angle cameras)
• JPEG compression artifacts

Mitigation:

• Use confidence threshold (70%+)
• Manual review for 70-89% range
• Request new photo if uncertain

Best Practices

1. Scan all documents - Even from trusted sources
2. Set 70%+ threshold - Balance false positives vs. fraud
3. Manual review uncertain cases - 70-89% confidence
4. Request new photo if flagged - Legitimate users can easily retake
5. Combine with liveness - Selfie + liveness prevents video replay
6. Monitor false positive rate - Adjust threshold if too many legitimate rejections
7. Keep models updated - New deepfake techniques emerge constantly

API Reference

Analyze Image

POST /api/v4/deepfake/analyse
 
// multipart/form-data
{
  "image": File,
  "analysisType": "document" | "selfie" | "generic",
  "returnDetails": boolean // Include full analysis details
}
 
// Returns
{
  "jobId": "job_abc123",
  "status": "processing" | "completed" | "failed"
}

Get Results

GET /api/v4/deepfake/jobs/:jobId
 
// Returns full analysis result

Batch Analysis

POST /api/v4/deepfake/batch
 
{
  "images": [
    { "url": "https://...", "type": "document" },
    { "url": "https://...", "type": "selfie" }
  ]
}
 
// Returns array of job IDs

Pricing

Next Steps