Photo of a Photo: How Certificall’s Sovereign AI Detects Fraud at 87%
A certified photo compliant with the eIDAS regulation carries recognized evidentiary weight under French and European law. Qualified timestamping, electronic seal, enhanced geolocation, SHA fingerprint, secure PDF/A-3B storage : the certification chain is robust by design.
But a certification chain cannot guarantee that what is submitted to it at the point of capture is authentic. This is true of every photo collection tool — mobile application, web portal, field solution. And that is precisely where some attempt to intervene.
The question Certificall has chosen to tackle head-on : how can you guarantee that the image submitted for certification truly reflects reality, and not an attempt at manipulation?
Manipulation attempts at the collection stage : a universal challenge
Every tool that collects photos as evidence is exposed to circumvention attempts. Certification secures the document after it has been created. It does not replace an analysis of the image itself.
Among the possible manipulations at the collection stage, the photo of a photo is one of the simplest to carry out. It consists of photographing an existing image — displayed on a screen or printed on paper — and submitting it as though it were a genuine capture taken in the field. No specialist tool required. A smartphone is enough. The attempt is discreet. And without a dedicated detection system, it is difficult to identify.
For businesses using certified photo solutions, the potential consequences are concrete:
| Business context | Manipulation risk |
|---|---|
| Insurance claim | Photos that do not correspond to the actual damage |
| Site inspection or property inventory | A previous or fabricated situation presented as current |
| Proof of service delivery | Attestation of work that was never carried out |
Most collection tools on the market have no mechanism to detect this type of manipulation. Certificall made the opposite choice — from day one.
Why general-purpose AI is not enough for this specific problem
When faced with this threat, the natural instinct is to rely on the most powerful AI models available. But on this precise task, raw power is not enough. And it raises a data sovereignty problem.
A general-purpose AI is trained to handle a very broad range of situations. It performs well across many contexts. But without training data drawn from the real-world context of field photo collection — device diversity, variable capture conditions, image types specific to professional use cases — its accuracy on photo-of-a-photo detection remains insufficient for professional deployment.
| General-purpose AI | Specialist AI (Certificall) | |
|---|---|---|
| Scope | Broad range of tasks | One task, one context |
| Training data | Generic, multi-domain | Drawn from Certificall’s real operational context |
| Sovereignty | Dependent on a third-party provider | Proprietary, hosted in-house |
| Accuracy (May 2026 benchmark) | ~50% on average | 87% |
This is the fundamental principle of an effective image and photo fraud detection AI : it must have learned to recognise the signals specific to the context in which it operates.
Contextual specialisation outperforms raw power.
Certificall built its AI on this principle. Not by adapting an existing model. By developing a proprietary, sovereign model, trained exclusively on this use case under the real operating conditions of the solution.
Certificall’s proprietary AI: sovereign, specialist, benchmarked
A model built from within
Certificall’s photo-of-a-photo detection AI is proprietary. It is not hosted by a third party, nor dependent on an external API. It is sovereign — developed, trained, and fully controlled in-house, using data drawn from the real operational context of the Certificall solution.
This is a structural choice. It guarantees Certificall’s technological independence, the confidentiality of processed data in compliance with the GDPR, and the ability to continuously improve the model without depending on an external provider.
The May 2026 benchmark
In May 2026, Certificall conducted a rigorous benchmark to objectively measure the performance of its AI against leading models on the market.
| Benchmark parameter | Value |
|---|---|
| Images analysed | 500 |
| AI models tested | 11 (including 10 general-purpose models) |
| Protocol | Identical prompt used for all AI models tested |
| Statistical measures | McNemar test + F1 score |
| Certificall overall accuracy | 87% |
| Ranking position | Equivalent to the global leader |
| General-purpose AI models outperformed | 8 out of 10 |
The McNemar test is a statistical tool that validates whether a performance difference between two models is genuine and not due to chance. A higher score alone is not enough: the difference must be statistically significant to be conclusive. Certificall’s results are presented on this basis.
A model trained on a single task, in its true business context, matches the most accurate general-purpose AI tested — and outperforms 8 out of 10 general-purpose models.
The Trust Score: from detection to decision
More than 30 criteria analysed automatically
The photo-of-a-photo detection AI feeds into the Certificall Trust Score — a confidence indicator calculated automatically for every dossier submitted to the solution, based on more than 30 criteria analysed simultaneously, grouped into 7 scores.
The Trust Score is an analysis layer that operates at the entry point of the process, before certification. It does not replace the client’s judgement: it provides a documented, timestamped, traceable indicator so they can validate with full information — not blindly.
Concrete value for B2B decision-makers
| Benefit | What it changes in practice |
|---|---|
| Fewer undetected manipulation attempts | Photo-of-a-photo attempts are intercepted before the dossier is validated |
| Better-documented decisions | A searchable, traceable analysis history for every dossier |
| Formalised accountability | A shared indicator between Certificall and its client organisations |
An eIDAS-compliant certified digital proof only carries its full legal weight if what it represents is authentic. The Trust Score is the mechanism that makes this authenticity verifiable at the point of entry.
A pioneer’s commitment, not a finished product
To our knowledge, Certificall is one of the first B2B certified photo solutions to have developed a proprietary, sovereign AI dedicated to photo-of-a-photo detection — validated by a rigorous statistical benchmark against leading models on the market.
87% accuracy is not 100%. Claiming otherwise would be a commercial posture, not a technical commitment. Transparency on this point is the hallmark of serious R&D work.
The objective is clear: enrich the training data, refine the model, continuously improve the detection rate — in step with the real-world usage of the solution. This is only one step.
eIDAS-compliant digital proof deserves protection commensurate with its legal value. This is a long-term commitment.
→ Contact Certificall to discover the Trust Score and see how our fraud detection AI integrates into your photo collection workflows.
Frequently asked questions
Is photo-of-a-photo fraud specific to certified photo solutions?- No. Photo-of-a-photo is a manipulation attempt that affects every photo collection tool, whether or not it certifies the images produced. Certification secures the document after it has been created — it does not replace an analysis of the authenticity of the image submitted at the point of entry.
- Why did Certificall build its own AI rather than using an existing model?
- General-purpose models, however powerful, do not have training data drawn from the real-world context of field photo collection. Certificall’s proprietary AI is trained exclusively on this use case, under the real operating conditions of the solution — which gives it superior accuracy on this precise task.
- What is the Certificall Trust Score?
- The Trust Score is a confidence indicator calculated automatically for every dossier submitted to Certificall. It is based on more than 30 criteria analysed simultaneously, grouped into 7 scores — including photo-of-a-photo detection. It allows users to validate a piece of evidence with full information, rather than blindly.
- What does sovereign AI mean?
- A sovereign AI is a model developed, trained, and fully controlled in-house, with no dependency on an external provider or API. Certificall’s AI is proprietary: the data it processes remains within the solution’s secure environment, in compliance with the GDPR.
Get in touch with us to explore our AI anti-fraud solution
