Fake receipts fool human reviewers but leave digital fingerprints that AI can detect. Here's exactly how document forgery detection works.
The Core Problem
Human reviewers see: A receipt that looks legitimate
AI sees: Compression artifacts, font inconsistencies, metadata anomalies, and layout geometry problems
When someone edits a receipt in Photoshop or a PDF editor, they leave traces in:
- Compression patterns (JPEG artifacts)
- Font rendering (anti-aliasing, spacing)
- Metadata (software signatures, timestamps)
- Layout geometry (alignment, spacing)
Let's break down each detection technique.
Technique #1: Compression Artifact Analysis
How It Works
Images go through compression when saved. Each compression cycle leaves mathematical patterns. Edited regions show different patterns than the original.
The process:
- Image divided into 8×8 pixel blocks
- Discrete Cosine Transform (DCT) applied to each block
- High-frequency components quantized (lossy compression)
- Edited regions go through this process twice
What we detect:
- Different quantization tables between regions
- Block boundary discontinuities
- Double JPEG compression signatures
Real-World Detection
Case: Photoshopped receipt amount
Original: $23.45
Edited to: $123.45
What DocVerify detects:
- Location: Amount region (x: 145, y: 89)
- Confidence: 0.94
- Signals: DCT coefficient distribution mismatch, Quantization table inconsistency, Double compression detected
- Original quality: 92, Edited region quality: 85
Technique #2: Font and Rendering Analysis
How It Works
Real receipts are printed by thermal printers or POS systems. Edited text is rendered by image editors or word processors. These leave different rendering signatures.
What we analyze:
- Anti-aliasing patterns - How character edges are smoothed
- Font weight consistency - Thickness of strokes across characters
- Character spacing - Kerning and tracking
- Baseline alignment - How text sits on invisible lines
Real-World Detection
Case: Fake hotel receipt from Word template
Claimed source: Thermal printer
Actual source: Microsoft Word → Print to PDF
What DocVerify detects:
- Text regions analyzed: 47
- Font clusters detected: 3 (expected: 1)
- Suspicious region: "$189.00"
- Antialiasing: Smooth gradients detected (computer-rendered)
- Conclusion: Text rendered by desktop software, not POS printer
Technique #3: Metadata Forensics
How It Works
Every file contains metadata about its creation, modification, and editing history. Legitimate receipts have predictable metadata signatures.
What we inspect:
- Creator/Producer software - What program created the file?
- Creation vs. modification dates - Time gaps indicate editing
- Edit history - PDFs store incremental save operations
- Camera/Scanner info - Real photos have EXIF data
Real-World Detection
Case: Bank statement edited in PDFescape
Claimed: "Original bank statement from Chase"
Reality: Edited in PDFescape.com to change balance
What DocVerify detects:
- Creator: Chase Online Banking
- Producer: PDFescape Web Editor v4.2 ← Red flag
- Created: 2026-01-15
- Modified: 2026-03-10 ← 54 days later
- Verdict: SUSPICIOUS - Document edited after creation
Technique #4: Vision Model Fraud Detection
How It Works
Beyond mathematical forensics, we use deep learning vision models trained specifically on document fraud.
What the model learns:
- Layout patterns of real vs. fake receipts
- Template signatures from fake receipt generators
- Contextual anomalies (impossible dates, bad math, inconsistent branding)
- Synthetic generation artifacts from AI-generated receipts
Training Data
The model is trained on:
- 500,000 legitimate receipts from real transactions
- 250,000 fake receipts (collected from fraud cases)
- 100,000 synthetically edited receipts (for augmentation)
Performance Metrics
Based on evaluation across 100,000 test documents:
| Metric | DocVerify | Human Reviewers |
|---|---|---|
| True Positive Rate | 96.3% | 68.2% |
| False Positive Rate | 2.1% | 12.4% |
| Detection Speed | 1.2s/doc | 45s/doc |
| Cost per 1000 docs | $8.90 | $450 (labor) |
What this means:
- DocVerify catches 96.3% of fake receipts
- Only 2.1% of real receipts get flagged (false alarms)
- 37x faster than manual review
- 50x cheaper than human reviewers
Get Started
Want to test these techniques on your own documents?
- Sign up - Get 10 free scans per month
- Upload a receipt - Test any image or PDF
- See the analysis - Get detailed forgery signals
Questions? Email technical@docverify.app