Fraud Detection 2026 – Trends in Theses  
Part 3: How does image analysis work?

When we examine digital documents with the human eye, manipulations often go unnoticed. And in any case, with today’s fully digitised application processes involving millions of pages, this would no longer be manageable. In Part 3 of our ongoing series on fraud detection, we therefore turn to image analysis. Because where a line has been shifted by a single pixel, a three-digit invoice amount may well have become a four-digit one…

So how can forged documents be exposed using state-of-the-art image-based fraud detection, also known as computer vision technology? Our trends and theses provide the answers:

What needs to be considered?

1. It almost sounds like archaeology: analysing compression artefacts. These artefacts are embedded in JPEG or PDF files and can reveal whether an image has been edited after the fact. Analysing compression artefacts is a key task in image analysis, as sudden changes in pixel structures often indicate tampering.

2. From digital archaeology to digital genetic engineering? In a sense, yes: so-called “cloned areas” (copy-paste of numbers or terms) may appear consistent to the human eye, but analytical tools detect divergent patterns.

3. Using a technique known as principal component analysis (PCA), colour deviations can be made visible, helping to identify subsequently inserted image elements or signatures.

4. Keyword: large-scale comparison. If humans alone had to review documents, it would take an enormous amount of time—and it would be nearly impossible to detect duplicates, layout deviations, or missing content. Who would realistically lay out 100 documents side by side to compare tiny details? For an automated solution like our document forensics tool ICO.Fraud, this is done within seconds.  
What else?

5. There are also limits to image analysis as a standalone tool for fraud detection. Printed and subsequently photographed forgeries require content-based checks or robust layout validation rules to be uncovered.

6. The art of modern fraud detection therefore lies in combining methods. That is why, in our product ICO.Fraud, we deliberately integrate a wide range of different verification techniques.

7. Individual anomalies, taken in isolation, can often be explained and do not necessarily justify suspicion of fraud. Only by combining multiple indicators in an intelligent scoring system can fraud cases be reliably identified and managed.

8. One final point: the human factor. But wait—only eight points this time, not ten? Exactly. Because this content is created by real experts in fraud detection, not AI tools trying to reach ten points at all costs simply because a prompt demanded it.

This brings us back, quite deliberately, to the importance of the human factor. It remains essential in fraud detection: only humans make the final decisions. Our fraud detection software provides a high-quality foundation of actionable insights.  
What does this mean overall?

Metadata analysis and image analysis are key components of state-of-the-art fraud detection. In the final part of this series, we will turn to “content-based analysis” and take a closer look at the EU AI Act and the collaboration between humans and machines when it comes to staying one step ahead of fraudsters with the support of software.

Image: prompted by ICO-LUX using Perplexity.ai