TrueSight is an innovative NLP and Vision-based redaction tool that allows user-defined gradational redaction, masking, and anonymization across multiple input formats, including text, images, PDFs, and videos. TrueSight is designed for high security, ensuring that no input data is stored or retrievable, making it an ideal solution for sensitive data.

• Gradational Redaction: Unlike traditional redaction tools, TrueSight allows users to define how much information is to be removed, enabling tailored redaction to specific needs.
• Multi-format Support: TrueSight supports multiple formats, such as text, images, PDFs, and videos, offering a comprehensive solution for various data types.
• End-to-End Privacy Assurance: TrueSight ensures that no input data is stored or retrievable, providing a high level of privacy and security for sensitive data processing.
• Customizable Output: Users can select the redaction level and tailor the output to meet their specific needs.

The PoC for TrueSight supports redaction and anonymization for:

• Text files
• PDF files
• Image files
• Video files

This table summarizes the performance of different models on a PII masking dataset. The custom DBERT model achieves the highest accuracy and F1 score, indicating its effectiveness in masking personally identifiable information (PII) compared to MobileBERT and DistilBERT.

To run TrueSight, ensure you have the following dependencies installed:

transformers
torch
streamlit
pandas
numpy
pymupdf
tf-keras
llama-cpp-python
ultralytics
opencv-python
torchvision

First download models in the streamlit folder Models - https://drive.google.com/drive/folders/1x5sv_F3odn013JD5-MZQqchJyeVtdNje?usp=sharing

To launch TrueSight's user interface, run the following command:

python -m streamlit run app.py