Banking and financial institutions need to have the correct and safe automated categorization of sensitive records, including invoices, tax forms, and bank statements. Standard rule-based systems and simplistic machine learning models are commonly unable to handle unstructured document layout, template variability, and Optical Character Recognition-induced noise, and therefore scale poorly and achieve low classification rates. To solve these problems, introduce a privacy-aware convolutional neural network architecture to classify financial documents based on text extracted by Optical Character Recognition. In the proposed pipeline, document images are converted to machine-readable text via OCR, followed by preprocessing steps, including token cleaning, normalization, and word embedding, to minimize recognition errors and noise. The convolutional neural network architecture is designed to identify contextual and semantic patterns in text sequences, enabling high-accuracy classification across a wide range of document types. To ensure compliance with financial data protection requirements, the framework incorporates mechanisms for data security, encrypted and controlled model storage, and a controlled inference process during training and deployment. The experimental findings indicate that the Secure convolutional neural network – Financial Document Classification is more accurate, efficient in processing, and more resistant to OCR noise than baseline classifiers. The proposed framework will facilitate scalable, secure, and automated financial document management across cloud and on-premises environments.