Videti še ne pomeni verjeti : zaznava ponarejenih osebnih dokumentov z velikimi jezikovnimi modeli

slovenian

2026

volume 93, issue 4

computer vision   detection of forged identity documents   explainable artificial intelligence   multimodal large language models   pattern recognition

Elektrotehniška zveza Slovenije
(Obvezni izvod spletne publikacije)

The rapid digitalisation of society has enabled remote identity verification based on photographs of identity documents, supporting a wide range of online services from banking to public administration. However, such systems remain vulnerable to document forgery and identity misuse, especially as modern image manipulation techniques become increasingly sophisticated and visually convincing. Traditional detection approaches rely on supervised models trained to recognize known artifacts and common forgery patterns. These methods struggle to generalize to novel manipulation techniques and typically provide only binary decisions without meaningful explanations, limiting trust in high-stakes applications. In this work, we investigate whether multimodal large language models (MLLMs) can address these limitations in a zero-shot setting. We present a large-scale systematic evaluation of six opensource models — DeepSeek-VL2-Tiny, LLaVA-1.6-Mistral7B, Qwen2.5-VL-7B, Gemma-3-12B, InternVL3.5-4B, and Pixtral-12B—on the IDNet benchmark, the largest standardised dataset of authentic and forged identity documents. Without task-specific training, we analyse their ability to detect forged documents and examine performance across different fraud categories and document templates. Our results indicate that MLLMs achieve strong performance in detecting diverse digital manipulations of identity documents. Nevertheless, a finer-grained analysis across issuing countries demonstrates that detection accuracy is significantly affected by the visual design and structural characteristics of individual documents.

20.10.2026