Watch Guard is a mini-project designed to detect and prevent access to nudity or explicit content on your device.
This Android application leverages a custom image classifier built with PyTorch. The classifier uses MobileNet V3 as its backbone, optimizing for both power efficiency and real-time inference on portable devices.
I am not experienced in Kotlin or Android development, hence I utilized GitHub Copilot (GPT-o3 & Sonnet 3.7 & Gemini Flash 2.0) to assist in building the Android application.
- The app takes a screenshot of your device every 10–20 seconds.
- The screenshot is processed through the classifier to check for explicit content.
- If explicit content is detected, the app automatically returns your device to the home screen.
- The process repeats continuously.
🚀 Early-stage development:
- Classifier Performance: Achieves 96% accuracy on the test set, but detection accuracy on mobile screens may initially be lower.
- App Performance: Optimizations are ongoing for efficiency and accuracy.
🔒 Privacy & Security:
- The app operates entirely offline, ensuring complete privacy.
- No data is transmitted or stored externally—your activity remains 100% private.
Please let me know if you would like to contribute to this app in any way and stay tuned for updates and improvements!