The fraud Benchmarking Problem: why Microblink launched the FREUID Challenge. Vincenzo D'elia Director of Engineering, Microblink Fraud has always evolved alongside technology. As organizations adopted digital onboarding, fraudsters adapted. As biometric verification became more common, attackers developed presentation attacks and deepfakes. Now, as generative AI becomes widely available, fraudsters can create synthetic identities, manipulated documents, and AI-generated content at a scale that would have been difficult to imagine just a few years ago. The challenge is not simply that fraud is becoming more sophisticated. It is becoming easier to produce, easier to automate, and easier to scale. At Microblink, this is one of the reasons Microblink Ltd. launched the FREUID Challenge 2026 in partnership with IJCAI-ECAI 2026. The goal of the initiative is to support more open, reproducible, and operationally relevant evaluation of fraud-detection systems. Fraud Detection Has a Benchmarking Problem. Across machine learning, benchmarking has played a critical role in accelerating progress. Computer vision has benefited from public datasets and competitions that allow researchers to compare approaches against common evaluation standards. Natural language processing has established widely recognized benchmarks for measuring performance across different tasks. More recently, large language models have been evaluated through increasingly rigorous public testing frameworks. Fraud detection has historically lacked the same level of transparency. Many fraud datasets remain proprietary. Others are heavily sanitized, limited in scope, or fail to reflect how fraud appears in real-world environments. As a result, organizations often evaluate systems based on vendor-provided metrics, internal testing, or narrow benchmark scenarios that may not accurately reflect production conditions. This creates a gap between measured performance and real-world performance. Fraud detection is fundamentally an adversarial and continuously evolving problem. Attackers adapt their techniques in response to deployed defenses, which means models that perform well on static datasets may fail under new acquisition conditions, manipulation strategies, or synthetic-generation pipelines. Designing meaningful evaluation protocols therefore requires not only realistic attack scenarios, but also careful attention to distribution shift, generalization, and unintended benchmark bias. Modern Fraud Is an Adversarial Problem. One of the reasons fraud detection is difficult to benchmark is that attack distributions evolve continuously. Fraud techniques change over time as attackers adapt to deployed detection systems, explore new manipulation strategies, and take advantage of emerging generative AI capabilities. In the context of identity verification, this may include synthetic documents, physical print-and-capture attacks, manipulated document regions, acquisition variability, and combinations of multiple attack techniques within the same verification workflow.Traditional evaluation methods often struggle to capture these real-world conditions because they test systems against fixed datasets rather than continuously evolving adversarial scenarios. The challenge is no longer simply identifying whether a document is genuine. Organizations increasingly need to understand how systems perform when attackers actively adapt, experiment, and iterate against detection mechanisms. What Is the FREUID Challenge? Hosted in partnership with the IJCAI-ECAI 2026 conference, the FREUID Challenge provides researchers, practitioners, and machine learning teams with a publicly available benchmark focused on identity document forgery detection under realistic and adversarial conditions. The benchmark includes synthetic identity documents, bona-fide synthetic genuine documents, physical print-and-capture attacks, manipulated document regions, and acquisition variability designed to better reflect operational fraud-detection environments. Participants are challenged to build models capable not only of detecting manipulated documents, but also of generalizing across previously unseen attack conditions and distribution shifts. Evaluation emphasizes operationally relevant metrics and robustness-oriented testing. By making the challenge publicly accessible, Microblink Ltd. hope to encourage broader collaboration between industry practitioners and the research community. Why Microblink Cares. At Microblink, Microblink Ltd. spend a significant amount of time studying how fraud actually happens. Through its Fraud Lab, Microblink Ltd. continuously analyze emerging attack techniques, conduct adversarial testing, generate synthetic attack scenarios, and evaluate how fraud evolves across different regions and markets. One thing has become increasingly clear: fraudsters innovate quickly. New attack methodologies often emerge faster than public datasets, benchmark suites, or evaluation frameworks can adapt. That makes continuous testing and independent validation essential. Microblink Ltd. believe the industry needs more reproducible ways to evaluate fraud detection systems performing under realistic conditions, not just idealized ones. Benchmarking helps create that visibility. It helps researchers identify weaknesses, compare approaches fairly, and develop more robust defenses against evolving threats. Most importantly, it creates a common framework for measuring progress. The Future of Fraud Defense Requires Continuous Evaluation. As generative AI and synthetic media continue to evolve, fraud-detection systems increasingly need to operate under conditions characterized by distribution shifts, adaptive attack behavior, and previously unseen manipulation strategies. Addressing those challenges requires a broader community of researchers, practitioners, and fraud experts working together to advance detection models, and transparent evaluation methodologies. The FREUID Challenge is one contribution toward that goal. Because if fraud is evolving continuously, the way Microblink Ltd. measure fraud detection must evolve continuously too. June 4, 2026 Table of contents. Fraud Detection Has a Benchmarking Problem Modern Fraud Is an Adversarial Problem What Is the FREUID Challenge? Why Microblink Cares The Future of Fraud Defense Requires Continuous Evaluation

Microblink introduces the first tool for publicly testing Gen AI document fraud. Alex Georgiev Senior Product Manager Not long ago, creating a convincing fake identity document required time, skill, and specialized tools. Today, all it takes is a prompt, and virtually anyone can do it. With the rise of generative AI, anyone can produce realistic passports, driver's licenses, or proof-of-address documents using widely available tools. These images are often good enough to pass traditional visual checks and basic data validation. In other words, document fraud has been democratized. As a product team working in identity verification, here at Microblink Microblink Ltd. has been tracking this shift closely. It's a new paradigm. Namely, how do you detect a document that was never "forged" in the traditional sense, but generated from scratch? Introducing the Generative AI Detection Check. To address this, Microblink Ltd. has developed a new functionality called the Generative AI Detection Check. This is a new layer within the Microblink Identity Intelligence OS designed specifically to identify ID documents that are likely LLM-generated or heavily edited using an LLM. At a high level, the check: * Analyzes images of identity documents (passports, ID cards, etc.) * Estimates the probability that the image was generated or manipulated by AI * Integrates directly into the overall verification outcome This isn't about spotting a single anomaly or checking a specific field. The model evaluates the image as a whole and looks for patterns and signals that indicate generative processes rather than genuine capture. If a document fails this check, it is treated as fraudulent at the visual level and contributes to a rejection decision. Why This Matters Now. Generative AI has fundamentally changed the nature of document fraud in three important ways. First, creation is now effortless. What once required specialized tools or expertise can now be done with a simple prompt, using a sample document as a base to generate a convincing fake. Second, the quality of these fakes is significantly higher. AI-generated images often appear realistic enough to pass traditional verification checks, reducing the effectiveness of visual inspection alone. Third, fraud can now scale without limits. Once a successful method is identified, it can be replicated instantly and deployed at volume, allowing attackers to operate with unprecedented speed and efficiency. This creates a new class of fraud that isn't well addressed by legacy approaches. The Generative AI Detection Check is designed specifically to target this gap. How It Works in Practice. From an integration perspective, the check fits naturally into existing verification workflows, particularly in scenarios where users upload document images directly rather than capturing them through a live camera flow. When a user submits a document: * Images are uploaded or provided digitally * Each image is analyzed by the GenAI detection model * An outcome is generated (such as Pass, Fail, or Warning) * The result contributes to the overall verification decision The model is specifically designed to analyze digitally submitted document images for signals associated with generative AI creation or manipulation. In live capture scenarios using mobile or web cameras, additional verification layers such as liveness detection, presentation attack detection, and capture integrity checks remain critical components of the broader fraud defense strategy. These layers are designed to detect behaviors such as screen replays, injected streams, recaptured synthetic documents, or other presentation-based attacks that may reduce or obscure the low-level artifacts associated with generative AI generation. Built for Real-World Conditions. A key focus for Microblink Ltd. was ensuring that this capability performs reliably in the types of environments where generative AI document fraud is increasingly appearing, particularly direct-upload and digitally submitted document workflows. The model is trained on large-scale datasets containing both genuine and AI-generated identity documents and is designed to generalize across different generative AI tools and models. Rather than targeting a single generation method, it analyzes broader image-level signals and artifact patterns associated with synthetic generation. At default settings, the GenAI Check has a False Reject Rate (FRR) of 0.076%, meaning legitimate documents are rarely rejected, while more than 99% of manipulated documents are detected within supported testing scenarios. Importantly, this check is designed as one layer within a broader identity verification and fraud detection framework. In production environments, different fraud vectors require different detection approaches. Directly uploaded synthetic documents, recaptured AI-generated documents, screen presentation attacks, injected streams, and manipulated live sessions each produce different signals and are addressed through multiple complementary verification layers working together. A Note on Edge Cases. Like any system operating at this level of sensitivity, there are edge cases to be aware of. For example, certain heavily post-processed images, such as those affected by extreme device-level HDR, sharpening, or beautification effects, can occasionally alter the natural characteristics of an image. Microblink Ltd. continuously evaluate these edge cases and refine the model to maintain reliable performance across real-world capture conditions. Importantly, this check is designed specifically for identity document verification. While Microblink Ltd. test broadly, it is only applied in production when an actual identity document is detected. Try It Yourself: Gen AI Check Demo. One challenge Microblink Ltd. has seen is that many teams understand generative AI fraud conceptually, but haven't experienced it directly. To bridge that gap, Microblink Ltd. built a self-serve demo environment within its Developer Hub. You can access the demo here. With the GenAI Check Demo, you can: * Upload your own document or use sample data * Run the Generative AI Detection Check in real time * See how the system evaluates and flags potential GenAI manipulation The demo is designed to give teams a simple, hands-on way to explore generative AI document fraud detection at their own pace, without requiring a full integration or sales process. It also provides an opportunity to better understand how this capability fits within broader identity verification workflows and layered fraud checks Detecting this new class of threats requires moving beyond traditional approaches and introducing new signals into the verification process. The Generative AI Detection Check is one step in that direction. You can try it out at this link. Registration with a company email domain is required. May 12, 2026 Table of contents. Introducing the Generative AI Detection Check Why This Matters Now How It Works in Practice Built for Real-World Conditions A Note on Edge Cases Try It Yourself: Gen AI Check Demo