IDSPACE: A Novel Document Generator for Reliable Evaluation of Digital Identity Verification Systems
Abstract
As services increasingly move online, trust institutions such as banks, lenders, and governments must verify the identity of remote users. While fraud detection tools are widely available, evaluating and fine-tuning them remains challenging because identity documents are sensitive and therefore scarce. Synthetic data generation offers a promising solution, and the demand is clear: our prior work in this area has been downloaded over 11,000 times (aggregated across eight parts). We introduce IDSpace, extending this line of research in three key directions. First, we propose model-guided Bayesian optimization, which tunes generation parameters to maximize both visual similarity and prediction consistency with target-domain models, using only a few samples from the target domain. Second, we decouple user-specified metadata (demographics, fraud patterns, capture device) from automatically tuned control parameters (font styles, noise levels, image quality), enabling users to configure evaluations without requiring low-level expertise. Third, we expand beyond template images to support scanned and mobile-captured documents. Experiments show that IDSpace improves evaluation consistency by 15β45% over baselines including CycleGAN, diffusion inpainting, and non-guided optimization, using only a few real samples, while improving training accuracy by up to 9% and SSIM similarity with the target domain by 10%. We also release a new dataset of 359,240 high-quality synthetic documents spanning ten European ID types.
1. Introduction and Motivation
The shift toward remote service delivery has amplified the need for robust digital identity verification. Financial institutions, government agencies, and other trust-bearing entities rely on automated systems to confirm that users are who they claim to be. However, the evaluation and fine-tuning of these fraud detection systems are hindered by two fundamental challenges:
- Data scarcity: Real identity documents are highly sensitive, making them difficult to collect, share, and use for research or benchmarking.
- Domain mismatch: Synthetic data often fails to capture the visual and statistical characteristics of real-world documents, leading to poor performance when systems are deployed.
To address these gaps, the research community has turned to synthetic data generation. Yet, existing approaches often require large amounts of target-domain data or lack the flexibility to model diverse user scenarios. In this context, we present IDSpace, a novel document generator that produces realistic identity documents for reliable evaluation of verification systems.
2. The Need for Synthetic Identity Documents
The exponential growth of online banking, e-government services, and remote onboarding has intensified the need for accurate identity verification. According to a 2026 market report, the digital identity verification market is projected to exceed $15 billion, reflecting the increasing reliance on automated checks for anti-money laundering (AML), know-your-customer (KYC) compliance, and fraud prevention.
Despite the proliferation of fraud detection tools, their performance is often evaluated on small or proprietary datasets, leading to inconsistent and non-reproducible results. A 2025 survey by the European Banking Authority highlighted that over 60% of financial institutions cite data scarcity as a primary barrier to adopting AI-driven verification systems. Synthetic data generation offers a controlled, scalable, and privacy-preserving alternative.
Our prior work in this domain (eight parts, collectively downloaded over 11,000 times) demonstrated the feasibility of generating realistic ID documents. IDSpace builds on this foundation, introducing innovations that make synthetic data generation more accessible and reliable for both researchers and practitioners.
3. IDSpace: Key Innovations
IDSpace addresses the limitations of previous synthetic document generators through three primary technical contributions.
3.1 Model-Guided Bayesian Optimization
Traditional synthetic data generation often uses fixed parameter sets or manual tuning, which is time-consuming and error-prone. We propose model-guided Bayesian optimization to automatically adjust generation parameters. Given only a handful of real target-domain samples, this method:
- Measures visual similarity (e.g., SSIM) between generated and real images.
- Measures prediction consistency by evaluating how well a target-domain verification model agrees on synthetic and real samples.
The optimizer iteratively proposes new parameter configurations that balance both objectives, converging to settings that yield high-fidelity, domain-relevant documents.
3.2 Decoupling Metadata from Control Parameters
A major usability barrier in synthetic data generation is the need for technical expertise. IDSpace separates:
- User-specified metadata: Demographics (age, gender, nationality), fraud patterns (e.g., alteration types), and capture device (mobile, scanner, webcam). These are intuitive for non-experts.
- Automatically tuned control parameters: Font styles, noise levels, image quality, and other low-level settings that significantly impact realism. These are optimized by the Bayesian approach.
This separation enables domain experts (e.g., fraud analysts) to design test scenarios without needing to understand the underlying image processing.
3.3 Multi-Format Document Support
Previous generators largely focused on template-based images. IDSpace extends support to:
- Scanned documents: simulated with layered noise, blur, and perspective distortions.
- Mobile-captured documents: mimicking camera angles, lighting variations, and reflective artifacts.
This capability ensures that verification systems are tested under realistic capture conditions, which is crucial for deployment in mobile-first environments.
4. Experiments and Results
We conducted comprehensive experiments to validate IDSpaceβs effectiveness across three metrics: evaluation consistency, training accuracy, and visual similarity (SSIM).
4.1 Experimental Setup
- Baselines: CycleGAN, diffusion-based inpainting, and non-guided optimization (e.g., grid search).
- Real samples: We used a small set (e.g., 10β50 images) from each target domain to guide generation.
- Dataset: Our released dataset contains 359,240 synthetic documents spanning ten European ID types, covering a range of security features and layouts.
4.2 Results
- Evaluation consistency: IDSpace improves consistency by 15β45% over baselines, meaning that verification scores are more stable across multiple test runs, reducing false confidence in system performance.
- Training accuracy: Models trained with IDSpace-generated data show up to 9% higher accuracy compared to those trained with baseline synthetic data.
- Visual similarity: Generated images achieve 10% higher SSIM with target-domain real images, confirming better perceptual fidelity.
These improvements demonstrate that IDSpace not only produces more realistic documents but also enables more reliable and reproducible evaluation outcomes.
5. The New Dataset
In addition to the generator, we release a companion dataset of high-quality synthetic IDs. This dataset is designed to support reproducible research in identity verification, covering diverse demographics and document types. Each sample includes associated metadata (e.g., demographics, fraud indicators) and mimics realistic capture conditions, making it suitable for benchmarking both traditional and deep-learning-based verification systems.
The dataset and code are publicly available to encourage adoption and further innovation in this critical area.
6. Conclusion and Impact
IDSpace represents a significant step forward in synthetic identity document generation. By combining model-guided Bayesian optimization, user-friendly metadata control, and multi-format document support, it addresses key obstacles in evaluating and training verification systems. The substantial improvements in consistency, accuracy, and realism, coupled with the release of a large-scale dataset, position IDSpace as a valuable resource for the computer vision and security communities.
As digital identity verification becomes more pervasive, the ability to rigorously and safely test these systems is paramount. We believe IDSpace will accelerate research and deployment, ultimately enhancing trust in online interactions. Future work will explore extending IDSpace to inter-document variability (e.g., physical degradation over time) and incorporating adversarial attacks to further stress-test verification algorithms.
This is an extended technical report submitted to arXiv on September 2, 2026. The authors are Lulu Xie, Yancheng Wang, Kanchan Chowdhury, Rolando Garcia, Yingzhen Yang, and Jia Zou.
via ArXiv CV