Valida: AI and NLP Techniques in Document Forgery Detection

This thesis delves into the development of a model for Valida, a system designed by Gradiant that detects and highlights modifications and alterations in documents. The objective is therefore to search for and identify within the documents the fields considered sensitive and most at risk of modification, namely codes, addresses, dates, and customer names. By leveraging advanced natural language processing (NLP) techniques, the proposed system combines OCR-based methods with cutting-edge multimodal large language models (MLLMs) to analyze document structures and extract relevant information. The thesis explores the integration of these models with parsers and disambiguators to enhance accuracy and address challenges posed by noisy data, diverse layouts, and multilingual content. A key focus of this work is evaluating the performance of OCR-free models, such as Qwen2-VL, and their potential as alternatives to traditional approaches. The study highlights the strengths and limitations of each method, including the adaptability of Qwen2-VL to unstructured inputs and the precision of DistilBERT in handling structured fields.