Banking & finance
Verify identity documents and detect forged statements during onboarding and review.
How the sector deploys itDeepFraudGuard
Check whether photographs, documents and identity papers are genuine, edited or AI-generated. Upload a file through the web app or send it from your own systems, and it returns a clear verdict — authentic, suspicious or forged — backed by visual evidence and plain-language reasoning anyone can follow
Product Walkthrough
Fraud ID Guard for identity documents
Aadhaar
A digit changed, and both systems said so
Both systems under DeepFraudGuard flag the submitted Aadhaar as forged. The automated checks name the inconsistency rather than only reporting one: the final digit has been modified, and the internal data anomalies that follow are what mark the document suspicious.
PAN card
Three edits on one PAN card
Both systems flag the submitted PAN as forged, and the automated checks name each alteration: the cardholder’s name and the father’s name were both retitled to “KUMAR”, and the PAN alphanumeric sequence was modified to match the standard formatting pattern.
Cheque
The MICR line was edited to agree
The specialised system that handles cheques and alternative documents flags this one as forged. The number in the top right corner has been modified — and the MICR code manipulated to stay consistent with it, so that the two would not contradict each other.
Receipt & invoice
One generated outright, one inconsistent
Two files with different tells. The receipt is flagged as forged on embedded digital markers showing it was artificially generated. The invoice is flagged for anomalies of its own — structural and logical inconsistencies inside its details.
What it is
Image, document & identity verification
Every file passes through several independent forensic signals at once — visual and frequency analysis, noise fingerprinting, synthetic-media signatures, provenance credentials, semantic visual reasoning and known-media matching.
Upload a file through the web app or send it from your own systems, and it returns a clear verdict — authentic, suspicious or forged — backed by visual evidence and plain-language reasoning anyone can follow.
Because a forgery has to fool all of them together, none slips through on a single weak check.
- Engines — image, document and identity — routed automatically
3
Engines — image, document and identity — routed automatically
- Independent forensic checks run together on every file
Multi-signal
Independent forensic checks run together on every file
- Per image in JPG, PNG, WEBP and more — plus PDF documents
20 MB
Per image in JPG, PNG, WEBP and more — plus PDF documents
- Key-authenticated, asynchronous, one endpoint for both types
REST API
Key-authenticated, asynchronous, one endpoint for both types
Try it here
Check an image or a document here
Drop an image or a document here
JPG, PNG, WEBP or PDF up to 20 MB
The engine is chosen for you · 1 credit
1 credit per check
Run the same engines the API runs — on a file of your own, with every signal family shown.
10 / 10Free checks left
See credit termsPreview of the output format. The signal scores and the verdict shown here are derived in the browser from the file’s name and size, not measured by the forensic engines — though the job flow, the engine routing and the corroboration rule are the real ones. Book a walkthrough to have files of your own run for real.
6 signal families are consulted on every file.
How it works
From a submitted file to an explainable verdict
01
Submit
Upload via the web app or REST API
02
Queue
Acknowledged instantly; processed in the background
03
Engine selection
Routed to the image, document or identity engine
04
Forensic signals
Multiple independent checks run together
05
Verdict + evidence
Authentic, suspicious or forged, with maps
06
Report
Downloadable result and plain-language reasoning
Product exclusiveness
What sets DeepFraudGuard apart
Layered forensic signals, not a single detector
Every file passes through several independent forensic signals at once — visual and frequency analysis, noise fingerprinting, synthetic-media signatures, provenance credentials, semantic visual reasoning and known-media matching. Because a forgery has to fool all of them together, none slips through on a single weak check.
Flagship capability · six signal families, run together
Deepfake & manipulated image detection
Flags AI-generated and digitally altered photographs and returns a clear REAL or FAKE verdict with a calibrated confidence score, so you can act on it with confidence.
REAL or FAKE · calibrated confidence
Document forgery detection
Detects tampering, splicing, inconsistent content and synthetic generation across documents and PDFs — invoices, statements, certificates, contracts — and pinpoints the exact tampered regions with an AUTHENTIC / SUSPICIOUS / FORGED verdict.
Authentic · suspicious · forged, with the region marked
Fraud ID Guard for identity documents
A purpose-built engine for Indian identity documents (Aadhaar and PAN) that validates secure codes, check digits, layout format and front-to-back consistency across every common Aadhaar layout as well as PAN — rules that can be checked rigorously rather than guessed.
Dedicated service · Aadhaar and PAN
Explainable results
Every verdict is backed by evidence — attention maps, frequency and noise views, error-level and recompression maps, tamper-region overlays with severity scoring, signal-contribution charts and a plain-language explanation that maps directly to what was found.
Evidence maps · signal contributions · plain language
Built-in false-positive guards
The document engine deliberately skips checks that raise false alarms on genuine papers, and a corroboration rule only escalates a verdict to near-certain when two independent signal families agree — keeping honest documents from being wrongly flagged.
Two signal families must agree to escalate
Continuous learning
A human-in-the-loop review process validates inputs, scores source reputation, recalibrates decision thresholds with a rollback safety net, and trains on difficult cases against a curated reference library — so accuracy improves over time.
Human-in-the-loop · rollback safety net
Developer REST API
Key-authenticated endpoints let your systems submit a file, receive a job identifier and poll for the result. Images and documents share one endpoint with automatic engine selection, so integration stays simple.
Submit, get a job id, poll for the result
How it compares
Measured against a single detector and a human expert
Detects AI-generated & edited images
- Single-model detector
- Partly — Single model
- Manual expert review
- Partly — Expert-dependent
- DeepFraudGuard
- Yes — Multi-signal forensics
Detects document tampering & forgery
- Single-model detector
- No — Usually image-only
- Manual expert review
- Yes — Yes, but slow
- DeepFraudGuard
- Yes — Images & PDFs
Verifies identity documents (Aadhaar / PAN)
- Single-model detector
- No — No
- Manual expert review
- Partly — Manual
- DeepFraudGuard
- Yes — Fraud ID Guard
Pinpoints the tampered region
- Single-model detector
- No — Score only
- Manual expert review
- Partly — Manual notes
- DeepFraudGuard
- Yes — Overlays with severity
Explains every verdict
- Single-model detector
- No — Confidence score only
- Manual expert review
- Partly — Subjective
- DeepFraudGuard
- Yes — Evidence + reasoning
Guards against false positives
- Single-model detector
- Partly — Varies
- Manual expert review
- Partly — Human error
- DeepFraudGuard
- Yes — Corroboration rule
Improves over time
- Single-model detector
- Partly — Periodic retrain
- Manual expert review
- No — No
- DeepFraudGuard
- Yes — Human-in-the-loop learning
Scales through an API
- Single-model detector
- Partly — Varies
- Manual expert review
- No — Manual throughput
- DeepFraudGuard
- Yes — Async REST API
The architecture
Six families in, one verdict out
One rule stands between them, and two have to agree
What goes in
One file · image or PDF
Routed automatically
- Image deepfake detection
- Document forensics
- Fraud ID Guard
- Visual & frequency
- Noise fingerprint
- Synthetic signatures
- Provenance
- Semantic reasoning
- Known-media match
Corroboration rule
One family is not enough. A verdict is only escalated to near-certain when two independent families agree — which is what keeps honest documents from being wrongly flagged.
Authentic · suspicious · forged
- Evidence mapsAttention · frequency · noise · tamper overlays
- Plain-language reasoningWhich families carried the verdict, and why
Use cases by sector
Use cases across sectors
Insurance
Spot manipulated photos and tampered documents in claims before payout.
How the sector deploys itLending & fintech
Validate income proofs, bank statements and ID cards at scale via API.
Government & public services
Authenticate submitted certificates and identity documents.
How the sector deploys itLegal & compliance
Check the authenticity of documentary evidence with an auditable trail.
How the sector deploys itE-commerce & marketplaces
Detect AI-generated or doctored product and listing images.
Deployment strategy
Submit a file, get a verdict — by app or by API
Submit a file
Upload through the authenticated web app or POST to the REST API. The upload is acknowledged immediately — images up to 20 MB in common formats (JPG, PNG, WEBP and more), plus PDF documents.
Automatic engine selection
The platform detects the content type and routes it to the right engine — image deepfake detection, document forensics, or identity verification — with no extra configuration.
Background analysis
Forensic analysis runs on background workers through a queue (PENDING → PROCESSING → COMPLETED), so throughput scales and the interface never blocks.
Retrieve the verdict
Track progress by job identifier and collect the verdict, confidence, evidence maps and plain-language reasoning as soon as they are ready.
Review, report & improve
Download reports, submit feedback on edge cases, and let the human-in-the-loop process recalibrate the models over time.
Under the hood
What the platform is made of
The delivery channels, the API and its limits, the architecture behind them, and who can do what.
- Delivery channels
- An authenticated web app for people who verify media by hand, plus a developer REST + JSON API over HTTPS for systems that need verification built in.
- API & limits
- X-API-Key authentication with a configurable daily rate limit per key, reset at midnight UTC. Two endpoints: one to submit a file and receive a job_id, one to poll for status and result.
- Architecture
- Asynchronous and queue-based. Uploads are acknowledged instantly, heavy forensic analysis runs on background workers, and clients track everything by a single job identifier.
- Roles & control
- Standard users upload, view history, download reports and give feedback. Administrators additionally manage users, learning and recalibration, the reference database and developer API keys.
Frequently Asked Questions
Three things, through three engines: image deepfake detection (REAL / FAKE), document forensics (AUTHENTIC / SUSPICIOUS / FORGED with tamper-region pinpointing), and identity-document verification. You submit a file and the platform detects the content type and routes it to the right engine automatically — no extra configuration.
It detects tampering, splicing, inconsistent content and synthetic generation in documents and PDFs, returning an AUTHENTIC / SUSPICIOUS / FORGED verdict and pinpointing the tamper regions. Evidence includes tamper-region overlays with severity scoring alongside the standard forensic maps.
It is a purpose-built engine for Indian identity documents — Aadhaar and PAN — that validates secure codes, check digits, layout format and front-to-back consistency. It is aimed at onboarding and review flows where you need to confirm an identity paper is legitimate.
Images up to 20 MB (JPG, PNG, WEBP and more) plus PDF documents. Upload through the authenticated web app or POST to the REST API. The API uses X-API-Key authentication with a configurable daily rate limit per key; images and documents share one endpoint with automatic engine selection. You receive a job identifier and poll for the verdict, which runs asynchronously on background workers (PENDING → PROCESSING → COMPLETED).
It has built-in false-positive guards that deliberately skip checks known to raise false alarms on genuine papers, and a corroboration rule that only escalates a verdict to near-certain when two independent signal families agree. A human-in-the-loop process continuously recalibrates the models, with a rollback safety net.
The rest of the line
More in SyntheticMediaGuard
Check an image, a document or an ID
Book a walkthrough and we will run your own files through the three engines, and show the evidence behind each verdict.