Talk to usTry the demo

Forged ID Document Classes and What a Photograph Can Reveal

Fraud and risk

For: Fraud or risk analyst who has to decide what a photograph of an identity document can and cannot prove

How forged ID documents are classed, which traces each class leaves in a phone photograph, and what public guidance and research say about detecting counterfeits, alterations, copies and recaptures.

You review identity documents that applicants photograph with a phone, and you need to know which kinds of forged document leave a trace in such a photograph, which leave none, and how much evidence exists that software can find the trace. This article is written for the fraud or risk analyst who has to draw that line. It starts from the classes of forgery that examiners and regulators describe, and then asks of each class what survives the trip from the physical object to a single image or a short video.

Everything stated as fact comes from the guidance and papers listed at the end, and each figure is attributed to the document that gives it. Where a paragraph is our own reading of that material, it says so. The article makes no statement about detection rates of any product.

Why classes matter

An examiner with a magnifier, an ultraviolet lamp and the document in hand answers a question about the object. A remote reviewer answers a question about an image of the object, and the image is a narrower channel. The United Kingdom Home Office National Document Fraud Unit states the aim of document verification as ensuring that the documents presented are both genuine and presented by the rightful holder. Those are two separate questions, and different forgery classes defeat them in different places.

A class matters because it predicts where the evidence is. A counterfeit shows its origin in the print and the material, an alteration shows it at the point of change, an impersonation shows it nowhere on the document, and a document rebuilt as an image shows it in the pixels. A review process that treats all of these as one thing called a fake will be strong against one class and blind to another.

The classes in the public guidance

The Home Office guidance distinguishes four types of document abuse.

  • Impersonation: the person is simply a look-alike presenting a genuine document.
  • Counterfeits: a complete reproduction from scratch to resemble an officially issued document.
  • Forgeries: a genuine document that has been unlawfully altered, for example by substituting a page, substituting a photograph or image, or altering the personal details.
  • Pseudo or fantasy documents: documents that may have the physical appearance of a passport or identity card but are not officially recognised and are not acceptable proof of nationality or identity.

The European Union Agency for Cybersecurity lists the attack methods that relate to the identity document in its report on remote identity proofing. They are an authentic document with one or more modified parts, such as a stolen or expired one; a complete reproduction of the document of a real identity; a complete production of a document for a fictional identity; a complete reproduction of a document with a partly real and partly fictional identity, for example with the date of birth changed; and a fantasy document created without reference to an existing type. The report adds that forgery can be physical, meaning changes made to the physical document, or digital, meaning changes made to the photo or video of the document.

That second axis is the one that matters for remote review. A physical forgery was made to survive inspection by a person holding the object. A digital forgery was made to survive inspection of an image, and it can be produced with no physical object at all. The rest of this article follows the chain from the object to the image, in the order in which a forgery can enter it.

Counterfeits: what the base material gives away

The Home Office guidance describes a counterfeiter as someone who will often simulate the security features that a reviewer would expect to see. It then lists the questions an examiner asks: what the overall quality of the document is, how it reacts under ultraviolet light, whether watermarks and other paper features are present, and whether it has been securely printed.

The examples in the guidance show that some of the traces are visible in ordinary light and some are not.

  • Print quality: background print on a genuine passport consists of solid fine lines, while the counterfeit shown uses random dots, which the guidance describes as a characteristic of readily available home printers.
  • Fibres: security fibres appear randomly across the paper, and in a counterfeit Greek passport the fibres on two pages sit in the same places, which the guidance says is not possible in genuine documents.
  • Watermarks: a genuine watermark has subtle variations in light and dark areas because the paper is thinner or thicker, while a counterfeit one has been printed onto the surface.
  • Holograms: the counterfeit example carries a generic hologram with the word GENUINE in it, in place of a detailed device that shows different colours and designs when rotated and tilted.
  • Raised print: genuine intaglio printing can be identified by touch, and the counterfeit example does not have raised print.
  • Ultraviolet reaction: counterfeit documents may fluoresce more brightly than genuine documents, which use dull, secure paper.

The guidance is also clear that this takes equipment and training. It says that magnifiers, ultraviolet light sources, transmitted light and document scanners are easy to obtain, and that the equipment is effective only if the user understands secure documents and their security features.

The European Union Agency for Cybersecurity makes a point that should lower any confidence in the idea that every counterfeit looks poor. It says that a significant percentage of documents have ineffective security features or none at all, that higher quality fake documents can have holograms, anti-tamper lines and perforation, and that detailed checks giving full assurance on the authenticity of every one of the thousands of acceptable document types remain impossible for any service provider.

What a photograph carries of a counterfeit

This section is our own reading of the sources above and not a statement from them.

  • Features that depend on touch cannot be in a photograph, and raised intaglio print is the example.
  • Features that depend on ultraviolet light or transmitted light cannot be seen in an image taken in ordinary light, and base fluorescence, the behaviour of fibres and the behaviour of watermarks fall here.
  • Features that depend on tilt appear only if the capture shows more than one angle.
  • Features that depend on the look of the print can be assessed from a sharp, well-lit image if the resolution is high enough, and the dots against solid lines in the guidance are an example of such a feature.
  • Cross-page consistency, such as fibres or patterns repeated in identical places, can be examined only if the capture includes the pages being compared.

The consequence is that counterfeits with good print can pass a visible-light photograph that an examiner with a lamp would reject. That is a property of the channel and not a flaw of any one method.

Forgeries: the trace sits at the point of change

A forged document starts as a genuine one, so the paper, the printing and the security features elsewhere on it are real. The Home Office guidance describes three common alterations: substitution of a page, usually the biographical data page; substitution of the photograph or image; and alteration of personal details.

Substituted pages

Passports are made from sheets of paper stitched together and cut to size, so the pages and cover should be in perfect alignment. To forge a passport it may first be taken apart and then reassembled by hand, which makes it difficult for all the pages to sit flush as before. The guidance also shows a forged Dutch passport whose machine readable zone contains incorrect characters, and it states that only the characters of the defined font should be seen in those lines. It adds a nuance that an automated system must respect: some countries have issued genuine passports with a non-compliant font in that zone.

Substituted photographs

Most photographs and images in passports and identity cards are protected by an ink stamp, an emboss or a laminate with visible safeguards. The guidance gives several signs of substitution.

  • An ink stamp over the photograph is produced with wet ink and should not be printed, and in the forged identity card shown the printed part of the stamp differs from the stamp on the rest of the card.
  • In a forged Portuguese passport, the new image was printed onto a piece of paper and placed on the page under the laminate, and the edges of that paper show in transmitted light while the paper fluoresces brightly under ultraviolet light because it is ordinary commercial paper.
  • Many forgers print a new image, or parts of an image, onto a thin clear laminate and place it on top of the page, and the guidance advises examining holographic and ultraviolet features because the new image may incorrectly sit above them.
  • Polycarbonate documents have a laser engraved image that should display shades of grey or brown, while forgeries usually have printed images containing jet black or coloured dots.

Altered details

The guidance notes that a forger often alters details together with a page or photograph. Its example is a passport whose date of birth was changed after the forger abraded the original pen ink and wrote a new date over it. The checklist for this class asks for damage to the paper around the details and around the photograph.

What a photograph carries of a forgery

This section is our own reading of the sources above and not a statement from them.

  • Some of the listed signs are structural and local: the misaligned or printed stamp, damage or a visible edge around the photograph, characters outside the defined font in the machine readable zone, a portrait built from coloured dots.
  • The signs that depend on ultraviolet or transmitted light, such as the bright fluorescence of added paper, are absent from an ordinary photograph.
  • A cross-check between the machine readable zone and the printed data does not need special light and is one of the most direct tests available from an image.

The United States Digital Identity Guidelines include such a comparison in their requirements. If a machine readable zone or barcode is present on the evidence, optical capture and inspection must compare the machine readable zone data to the printed data on the evidence for consistency.

Impersonation and genuine documents in the wrong hands

In an impersonation the document is real, and a document forensics method has nothing to find. The guidance proposes the checks that apply: whether the person can recreate the signature without sight of the document, whether the person looks the correct age, whether distinguishing features such as moles or scars match, and a feature by feature comparison of the face. In remote review the equivalent is the comparison of the selfie with the portrait, together with a check that a live person is present.

The European Union Agency for Cybersecurity adds that authentic documents can be used to carry out an attack, so it is essential to verify that a document is not lost, stolen or expired. It names the stolen and lost travel documents database, SLTD, as one place to look, and its footnote identifies that database as INTERPOL's. It notes that the check only mitigates the risk and cannot rule it out, because a lost or stolen document might not be reported as such. INTERPOL describes the database as holding information on travel and identity documents that have been reported as stolen, lost, revoked, invalid or stolen blank. Only the country which issued a document can add it to the database, and law enforcement officials reach it through INTERPOL's secure police communication system. For the private sector INTERPOL describes a separate initiative, I-Checkit, in which trusted partners submit travel documents for screening when customers book a ticket, and a positive hit is relayed to law enforcement. A verification service should therefore not assume that it can query the database directly. For documents with a chip, it describes the data as authenticated through public key infrastructure, which makes forgery borderline impossible with current technologies when all security mechanisms are fully and correctly implemented. It also records the limit of that route: chip readers are not available on older and low-tier smartphones.

A face-morphing attack sits between the classes. The portrait on the document is a blend of two faces, so that the document can be used by more than one person. The IDNet dataset includes face morphing and portrait substitution among its six forgery patterns, and the paper notes that a dataset which blurs the portraits is unsuitable for portrait substitution and face morphing work.

Copies and recaptures: when the object is not the document

A different family of attack does not alter or counterfeit the document at all. The applicant photographs a copy, a printout, or another screen showing an image of the document. Research datasets model these cases as recapture attacks and as presentation attacks on a document with holograms.

The Document Liveness Challenge dataset, DLC-2021, takes mock documents as the originals and films three attack types: a colour printed copy without lamination, a grey printed copy without lamination, and an image of the document displayed on a screen. The authors describe it as a basis for evaluating document image recapture detection, photocopy detection and lamination detection. Its counts by type are 290 laminated originals, 484 unlaminated colour copies, 250 unlaminated grey copies and 400 screen recaptures.

What the paper says about the visual evidence is useful for a reviewer. Laminated documents easily introduce reflection and saturation, especially when a strong light such as a flash, a fluorescent lamp or the sun lights the document during capture. The authors write that strong reflections can partially or totally hide the content, and that their shape and size vary with the orientation of the document relative to the lens. They also write that the shape and consistency of changes in highlights and in scene geometry can serve as an important indicator of the liveness of a document. In their baseline for unlaminated colour copies, they note that the presence of glare is the most evident feature of laminated documents.

The same paper warns about false alarms. Documents themselves may have a complex textured page background, for example from guilloche, and textured objects or even a screen behind the document can produce moire and other recapture artefacts outside the document zone when the original is captured with a digital camera. Partial occlusion of a grey copy also raises the colour diversity of the document area, which makes grey copy detection harder.

The baseline results are modest, which is informative in its own right. For screen recapture detection the authors trained a classifier based on ResNet-50 on patches, and for grey copy detection all the convolutional models they tried either did not train at all or overfitted, so only a trivial reference classifier is reported.

Holograms in video

The MIDV-Holo dataset contains 700 video clips of identity card-like and passport-like documents with a hologram. For each fictitious identity there are three original clips and four presentation attack clips. The attacks are a copy without a hologram, a pseudo hologram copy made as a static imitation in an image editor, a photocopy of the document, and a photo replacement in which an original document is physically altered to change the face picture. In the photo replacement attack no hologram is visible over the face although it is still present on the rest of the document.

The paper that works with this dataset describes three controls and what each can catch.

  • Appearance conformity asks whether a visual appearance can be generated by a genuine optically variable device, and it detects attacks with no hologram or a different shape, but would be tricked by simple static copies of the expected hologram.
  • Appearance coverage asks how well a set of appearances, usually from a video, covers the set of possible ones, and it detects static holograms, but would be tricked by any random holographic layer if used without the first control.
  • Transitions validity asks whether the changes between observed appearances are consistent with the expected ones, and it detects imperfect imitations or rapid swapping of static holograms.

The authors add that the low frame rates in current real remote authentication applications are a significant challenge that has not been explored adequately in the literature. The practical reading is that a single still image of a hologram carries little, a tilt video carries more, and the quality of the video bounds what the controls can do.

The European Union Agency for Cybersecurity reaches a similar conclusion at the process level. For high-level verification of a document without a chip, it says the verification should require a video of both the front and back side while the document is tilted to check variable optical features. It says that video based solutions provide more data for analysis and therefore higher assurance than photographs, and that the verification needs both specialist automatic means and human verification by an expert operator.

Digital alterations of the image

A digital alteration changes the image after, or instead of, the capture. The authors of the Synthetic Dataset of ID and Travel Documents, SIDTD, note that forged documents are scarce compared with genuine ones and that the way they are produced varies from fraudster to fraudster, which gives the forged class a high internal variability. They generate forged versions with two techniques.

  • Crop and Replace exchanges a region between two documents of the same type, with a random shift so that the match is never perfect and a texture discontinuity appears at the border.
  • Inpainting removes a text field, rebuilds the background and writes new content in a chosen font.

The authors observe that, depending on the shift and the font, the result can range from easily distinguishable by a person to extremely subtle.

IDNet extends the set of patterns to six: crop and move, inpaint and rewrite, text replacement, face morphing, portrait replacement and a mixed pattern. Its authors say the choice of patterns is informed by their prevalence in real-world fraud and by their overlap with personal identifier fields. They report that a face morphing detector trained on one US driver's licence type dropped from 98.65% to 50.55% accuracy on another, and they conclude that fraud detection models rely heavily on type-specific features such as background patterns and size. For a reviewer the lesson is not about one model. It is that a detector which learns the look of a document type can look excellent on that type and fail on the next, and that a vendor's claim should state which types it was measured on.

The paper adds that most remote identity validation services rely on images captured in white light and not on multi-spectral imaging, and that this is the scenario the dataset addresses.

The European Union Agency for Cybersecurity notes a harder case, in which digital tools to make fake identity documents are available, and software that uses deepfake technology makes the identification of changes on a human visual basis practically impossible and calls for controls of another nature. It also describes software that can apply the security features typical of a specific document to a simple piece of paper by injecting an overlay in real-time video, complete with reflections and the correct management of overlaps. That attack belongs with video injection and not with document analysis, and the guidelines treat it separately.

Documents generated by image models

A 2026 study asks whether publicly available diffusion models can produce identity document forgeries that could bypass human or automated verification. The author tested text-to-image and image-to-image pipelines from multiple model families and concluded that current generative models can simulate surface-level aesthetics but fail to reproduce structural and forensic authenticity. The study describes itself as a qualitative, expert-driven forensic analysis and not a quantitative benchmark, and it is the work of one author, so it is evidence of direction and not a measurement.

The failure patterns it lists across the models are practical cues for a reviewer.

  • Text errors: unreadable, distorted or nonsensical characters that resemble pseudo-text.
  • Unnatural uniformity and lack of wear: documents that look overly pristine, flat or digitally edited, without the subtle imperfections of real documents.
  • Degraded security elements: overly smooth gradients, absence of micro-relief, and uniform shading, which resemble a digital redrawing of the element.
  • Letter deformation and loss of fine patterns, including microtext that becomes unreadable or visibly distorted.
  • Incorrect document scale compared with real plastic card dimensions, with implausible proportions and perspective.
  • Portrait imitation: an imitation of a laser engraved portrait showed a grayscale digital photo appearance, artificial grain and uniform surface texture, and an imitation of an inkjet portrait showed no observable halftone or micro-dot structure.
  • Collateral changes: in portrait substitution the models changed texture and text beyond the intended region, and flattened security patterns that extend across the background and the portrait area.

The study also looks at services that sell images of fake documents. From the public samples it analysed, the author finds that such services appear to rely on fixed document templates and traditional image-editing workflows, and says that it remains unclear how far generative models are used as a core component. The author draws the conclusion that the risk of generative documents achieving forensic-level authenticity may be overestimated, and recommends that machine learning practitioners work with document forensics experts.

A study of this kind describes a set of models on the date of the experiment, and models change quickly.

What the guidelines ask of image-based validation

The United States Digital Identity Guidelines treat the validation of physical evidence as either optical capture and inspection, often called document authentication, or visual inspection by a trained proofing agent, and allow a provider to use either or both. For optical capture and inspection, the automated technology must meet a document false acceptance rate of 0.1 or less and a document false rejection rate of 0.1 or less. The guidelines require live capture of documents during validation together with document presence checks, and they explain that live capture techniques confirm that the document is physically present and that the image captured is not a manipulated digital copy.

Two requirements bear directly on how a result should be read. The provider must assess the technology in conditions that are substantially similar to the operational environment and the types of evidence presented by the user base, covering all the evidence types it allows to be validated this way. And it must make the results of its testing publicly available.

For proofing agents, the training must cover authentic layouts, physical security features, techniques for assessing features, including where tactile inspection is needed, and common indications of tampering such as damage to the lamination or image modification. The guidelines say they do not attempt to provide a comprehensive list of security features, so providers need training specific to the types of evidence they accept. Agents who inspect by remote means must be provided with devices and connections that support sufficiently high-quality imagery.

Public registers of genuine documents support both sides. An archived case study on the European register of authentic identity and travel documents, from 2008, describes it as a source of information on the security features of authentic documents for the general public, including employers, banks, credit institutions and security companies. It explains that the register is generated daily as static pages from a system that holds information on authentic and false documents and that only a small group of document experts in the Member States can access that system. The European Union Agency for Cybersecurity also names such registers as very useful to providers who have to maintain a database of acceptable document types.

A class by class summary

This summary is our own reading of the sources above.

  • Counterfeit with poor print: likely to show in a sharp visible-light image through the print structure, layout and fonts.
  • Counterfeit with good print: may need ultraviolet, transmitted light, touch or the chip, which a photograph does not provide.
  • Page or photo substitution: leaves local traces such as edges, damage, a printed stamp or a mismatch in the machine readable zone, which a sharp image can show if the area is not covered by glare.
  • Altered details: traces are small and local, and the consistency between the printed text and the machine readable zone is an inexpensive test.
  • Copy, printout or screen recapture: leaves traces in the capture such as glare behaviour, moire, missing optically variable behaviour and print texture, and video helps more than a still image.
  • Digital edit of the image: leaves traces in the pixels, and detectors tend to learn the document type they were trained on.
  • Generated document: tends to show uniformity, flattened security elements and wrong text, according to one qualitative study.
  • Impersonation: leaves nothing on a genuine document, and the evidence is in the face comparison and the liveness check.

Questions to ask of a document check

  • Which of the classes above was the result measured on, and were the forgeries physical, printed copies, screen recaptures or digital edits?
  • Which document types were in the test, and does the result hold on a type that was not used to train the model?
  • Were the images captured on the devices and in the lighting of the real user base?
  • Does the check use a single still image or a short video, and what happens to the result when glare covers the photograph area?
  • Does the check compare the machine readable zone with the printed data, and how does it handle genuine documents with a non-compliant font?
  • What are the false acceptance and false rejection rates for documents, and is the test public?
  • Which classes are routed to a person or to another step, and how are lost, stolen and expired documents handled?

For the related requirements on measuring and reporting these rates, see the companion article on automated document validation.

Where this sits in a verification flow

The document step of a NeuroVision verification session detects, classifies and reads IDs. It reads printed and handwritten text and the machine readable zone with its checksum, and it carries signature and stamp checks. Document anti-fraud covers screen photos, photocopies, printouts, glued photos, drawn-on elements, erasures, edited images and logical data errors, and it is included in every package. The classes above are a way to ask which of these checks matter for your document mix and which classes need another step. To see how these steps look in your own scenario, contact the team.

Sources

All sources were fetched on 27 October 2026.

Ready to talk?

Tell us what you are building and we will help you plan the integration.

Talk to us