The Outlier Lab

Receipt and Document Image Forensics

Signals extracted from submitted receipts, checks and documents (structural consistency, merchant template conformance, duplicate and near-duplicate matching, metadata and synthetic-image indicators) and their limits.

What it measures

Whenever a platform pays money because of an image (cashback on a receipt, availability on a mobile check deposit, a refund on a "damaged item" photo, a rebate on an invoice), the image is effectively the transaction. Image forensics asks five layered questions:

Diagram: Submitted image, 1. Capture: real photo of a real object?, 2. Structure: internally consistent?, 3. Template: matches how this issuer formats documents?, 4. Uniqueness: seen before, here or elsewhere?, 5. External truth: does independent data confirm it?, Pay, hold, review, deny
LayerWhat it examines (category level)
CaptureSigns the image is a photo of a physical document vs. a screen capture, re-photograph, scan of a print, digitally composed or synthetic image; capture metadata (device, time, software) where present; consistency of lighting, perspective, paper texture and focus
StructureArithmetic integrity (line items, discounts, tax and totals agree); internal date/time logic; tender lines consistent with totals; text rendering consistency (fonts, spacing, alignment) across the document
TemplateConformance to the merchant's or bank's known formats: layout, header and footer structure, identifier formats, tax treatment for the store's jurisdiction, item description conventions; drift from that template between submissions
UniquenessExact duplicates (cryptographic hash), near-duplicates (perceptual hash or learned image embeddings), and semantic duplicates (the same transaction identifiers, totals and timestamps appearing on visually different images)
External truthDoes the store exist and was it open; are prices plausible for that retailer and date; does a linked card or loyalty record show a matching transaction; for checks, does the issuing bank or check-verification data recognize the item

Why it matters

In Cashback, Coupons and Rewards, receipt submission is the entire value-entry point: the platform pays based on what the image claims. The same is true for remote deposit capture in Neobanks and Consumer Fintech, for proof-of-purchase in rebate programs, and for photo evidence in Online Marketplaces claims and Retail and Ecommerce returns.

What changed in 2025–2026 is the cost of producing a convincing fake. AppZen reported that AI-generated images rose from none of its flagged fake receipts in March 2025 to about 71% by mid-May 2026, overtaking template-based fakes in April 2026; it counted 1,471 AI-generated receipts from 745 employees at 174 companies. SAP Concur said it detected 18 times more suspected AI-generated receipts after deploying an AI receipt checker, with about 1% of receipts flagged. Those are expense-management numbers, but cashback and rebate programs face the same generators with a stronger incentive: no employer, no audit, and payouts that scale with account count. AppZen also found AI fakes were smaller on average (about $100, median about $32) than template fakes, small enough to slip under value-based review thresholds.

Checks have their own document-forensic tradition. FinCEN's mail-theft check fraud alert (FIN-2023-Alert003) lists check stock noticeably different from the issuing bank's and faded handwriting under darker overwriting as red flags. Those are exactly the alterations and counterfeits that RDC image analysis aims to catch.

How to collect it

  • Keep the original bytes. Store the file as uploaded (with any metadata) plus your processed version. Many pipelines re-encode on upload and destroy evidence.
  • Record capture context: in-app camera vs. gallery upload, device and session (Device Fingerprinting), time from capture to submission, number of retakes. In-app capture with liveness-style checks (multiple frames, edge detection of a physical document) raises assurance; gallery uploads lower it.
  • Run OCR and structured extraction into fields (merchant, store ID, timestamp, items, totals, tender, identifiers) so structure, template and semantic-duplicate checks can run.
  • Maintain hash indices of exact hashes and perceptual hashes/embeddings for every submission, searchable across all accounts and time. Duplicate detection gets more valuable as the corpus grows.
  • Keep a merchant template library: learned or curated representations of how high-volume merchants format receipts, versioned as merchants change POS systems.
  • Bring in external data: card-linked transaction data, retailer or POS partnerships, e-receipt integrations, and check-verification and duplicate-presentment services for checks.
  • On build vs. buy, vendors offer document forensics, synthetic-image detection and receipt OCR as APIs; RDC vendors bundle image quality and duplicate checks. Your cross-submission corpus, merchant templates for your offer mix and links to accounts are things only you can build. Most mature programs combine both and benchmark vendors on their own labeled fakes.

How to use it

  • Treat forensic output as a risk score. Route by score and value: auto-approve clean, low-value submissions from trusted accounts; hold or vest higher-risk payouts (Reward Vesting and Clawback); send ambiguous cases to Manual Review Queues with the specific anomalies highlighted.
  • Combine with Basket and SKU Anomalies: a structurally perfect receipt whose basket is implausible (only offer-qualifying items, unusual quantities) is a stronger signal than either alone.
  • Combine with linkage: near-duplicate or template-identical receipts across different accounts reveal shared-receipt rings and Multi-Accounting (Entity Link Analysis). The same receipt image appearing under many accounts is often the first visible sign of a farm.
  • Combine with Promo and Reward Redemption Patterns: receipts clustered around a newly launched high-value offer from new accounts deserve tighter thresholds.
  • Shift weight to external truth over time. As generated images improve, pixel-level detection will keep degrading; independent confirmation (card transaction match, retailer data, e-receipts) does not depend on how good the image looks. The strongest receipt programs reduce how much they have to trust the image at all (Receipt and Purchase Verification).
  • For RDC: duplicate-presentment checks (same item deposited twice, here or at another institution), check stock and MICR consistency, and alteration indicators feed Funds Availability Holds.

Strengths and limitations

Strengths: acts at the point of value entry; duplicate matching is cheap and highly precise for reuse; structural checks catch careless fakes at scale; evidence is visual and easy for reviewers and QA to evaluate.

Limitations:

  • Perceptual hashes can be attacked. Academic work has shown that both widely used classical perceptual hashes and deep learned hashes can be deliberately manipulated to avoid matches or to force false matches (Dolhansky and Canton Ferrer, 2020; Struppek et al., 2022; black-box results in 2026). Use them for near-duplicate triage, combine multiple hashing approaches and semantic field matching, and don't treat "no match" as proof of uniqueness.
  • Synthetic-image detectors generalize poorly to generators they weren't trained on, and detection accuracy is a moving target. Expect periodic sharp drops in recall when new generation models are released.
  • Provenance metadata is incomplete. Standards such as C2PA Content Credentials and generator watermarks help when present, but support is inconsistent across generators, signed credentials attest to claims rather than reality, and an image without credentials tells you nothing definitive. Metadata of any kind is frequently missing for benign reasons and can be lost or altered, so its absence is weak evidence.
  • Templates drift. Merchants change POS systems and receipt layouts; stale templates generate false positives.
  • Any of these signals can be degraded by a determined, well-resourced adversary. Compensate with external transaction confirmation, payout vesting, account-level and network-level signals, and offer design that limits exposure per receipt.

False positives and fairness

  • Crumpled, faded, thermal-paper receipts and poor lighting, all common for legitimate shoppers, especially on older phones.
  • Legitimate e-receipts and screenshots from retailer apps, which look "digital" by design.
  • Small independent stores with non-standard or handwritten receipts, and merchants that recently changed POS formats.
  • Households legitimately submitting the same receipt to different programs (allowed on some platforms and not others, so define it in the terms).
  • Extreme couponers with very high submission volume and highly optimized baskets.
  • Non-English or multilingual receipts that OCR handles poorly.

Measure false positive rates by merchant, device tier and language, and give customers a resubmission path before denial.

Privacy and compliance considerations

Receipts can contain partial card numbers, names, loyalty IDs, addresses and pharmacy items (potentially health-related); minimize what you extract and retain, redact in reviewer tools, and set retention to dispute and audit needs. If you use submissions to train models, disclose it. For RDC, check images are subject to banking record-retention requirements and Reg CC hold rules. Revoking earned rewards should follow clearly disclosed terms (Privacy and Data Use in Fraud Prevention).

Mechanisms it helps detect

Receipt Fraud (Cashback) · Check and Mobile Deposit Fraud · Generative AI and Deepfake Fraud · Refund and Return Fraud · Buyer Protection Claim Abuse (INR and SNAD) · Coupon Counterfeiting and Misredemption · Multi-Accounting · Fraud Rings and Collusion

Basket and SKU Anomalies · Promo and Reward Redemption Patterns · Entity Link Analysis · Device Reputation and Linkage · Receipt and Purchase Verification · Document Verification and Liveness · Reward Vesting and Clawback · Playbook - Receipt and Offer Fraud Members · Playbook - Mobile Deposit Fraud Members · Playbook - Novel Attack Detection Members

Sources and further reading