Doc-Patch: An Unsupervised Approach for Documents Forgery Detection

The exponential growth of digital documents has unfortunately led to a parallel rise in document forgery. Traditional authentication methods often struggle to detect “unseen-before” sophisticated falsification techniques. To address this challenge, we introduce Doc-patch, an unsupervised approach that effectively identifies anomalies in documents without requiring labeled training data. Doc-patch leverages advanced machine learning techniques to analyze document patches, focusing on small segments of text or images. By employing feature extraction, self-attention layers, and subtle hint capturing layers, the model can pinpoint local anomalies indicative of tampering. The FAISS K-Nearest Neighbors algorithm is then used to identify and localize these inconsistencies. Our experimental results demonstrate that Doc-patch achieves a high average precision score of 96.51% on a diverse dataset of documents. This superior performance underscores the effectiveness of our approach in detecting document forgery, surpassing traditional methods and providing a robust solution for authenticating digital documents in professional and administrative settings.