This paper presents a neural network–based approach for automated detection of forged areas in PDF documents. The method converts document pages into image representations and applies semantic segmentation to localize manipulations such as inserted signatures or replaced text. U-Net and DeepLabV3+ architectures were trained on a dataset of authentic and synthetically forged samples. Experimental results show that segmentation models achieve high accuracy (IoU ≈ 0.91), outperforming traditional metadata-based techniques. The proposed solution can be integrated into electronic document management and cybersecurity systems for automated authenticity verification.