Security filtering solution
The business goal
The business goal of the project was to create a tool that filters unwanted content. This solution works as a client-server application. This application involves using Computer Vision techniques in order to find and filter unwanted or abusive content.
From the system programming side the app is a cloud control and managed software agent that provides centrally managed security for devices running Microsoft Windows. The application consists of several main components:
- Management Service is the main control service performing high level coordination and supervision leaving the heavy lifting and performance critical work to the low level libraries or services.
- Redirector is a component that filters network traffic and redirects it to the cloud proxy service if necessary, providing internet security according to the desired use policy.
- The watchdog is a component that monitors the health of the Agent, provides automated self healing, and tamper detection and prevention.
From the Data Science perspective we researched different existing models skin and nudity detection. We analyzed and improved the model, and made it faster and more accurate. The ultimate goal was to make an original model as fast as possible (as close to real-time as possible) and improve the accuracy in order to decrease the amount of false detections.
Since we applied a Deep Learning model, the system can be improved by retraining models using new samples of nudity/pornography images by fine-tuning the original model for custom requirements.
Ultimately, this project resulted in a robust solution that could be used by the users. This solution proved its effectiveness and eventually providing feedback for later fine tuning, like adding a service for self pre-moderation and learning of the content.