🛡️Image Interceptor : AI-Powered Real-Time Harmful Image Filtering Browser Extension

Project Overview

Image Interceptor is an AI-powered real-time harmful image filtering program designed to automatically detect, block, and replace unsafe images on websites and social media platforms.

✔ AI scans all images on active webpages
✔ ConvNeXt-powered harmfulness classification
✔ Instant replacement with safe placeholder (logo)
✔ Works directly as a browser extension

---

Why Image Interceptor?

Issue	Limitation in Existing Solutions	Image Interceptor Advantage
Users exposed to harmful visual content (NSFW, violence)	URL-based filtering cannot detect images	Pixel-level harmfulness judgment by AI
More graphical / mixed-content images	Rule-based logic easily fails	Context-aware filtering
Hard to use security features	Ordinary users cannot configure	One-click activation, fully automated

---

AI Model

Dataset

• ~70,000 images collected, including:

  • Safe: daily-life images representing real browsing context
  • NSFW: nude / violent imagery
  • Additional false-positive mitigation data (e.g., fashion, art)

• Image processing:

  • 224×224 resize
  • RGB normalization
  • ColorJitter, RandomCrop, HorizontalFlip

Model Architecture (ConvNeXt-Tiny)

Input (224x224 RGB)
      ↓
ConvNeXt-Tiny Backbone
      ↓
Global Average Pooling
      ↓
Fully Connected Layer
      ↓
Softmax (Safe / NSFW)

---

Performance Summary

Model Version	Feature	Accuracy	Precision	Recall	F1
v1 Baseline	Safe + NSFW only	0.91	0.895	0.902	0.890
v2 Improved	+ false-positive protection data	0.933	0.948	0.920	0.934

➡ Significant improvement across metrics
➡ Lower misclassification of fashion/art images

---

System Architecture

[Browser Extension]
   ↓ scans <img> tags
[Image Collector → Base64 Encode]
   ↓
[FastAPI Server]
   ↓
[ConvNeXt AI Model]
   ↓
Classification Result
   ↳ Safe  → Display original
   ↳ NSFW → Replace with logo

⏱ Real-time response while browsing

---

Visual Outcomes

Case	Original	Prediction	Output
Nude content	NSFW	❌ Replaced
Violence scene	NSFW	❌ Replaced
Normal daily image	Safe	✔ Retained

➡ Protects users without ruining their browsing experience

---

Technologies

Component	Stack
AI Model	PyTorch, ConvNeXt-Tiny
Backend API	FastAPI + Uvicorn
Browser Extension	JavaScript (DOM & Web APIs)
Deployment	Google Cloud Platform

---

Expected Impact

• Youth protection
• Digital wellness enhancement
• Strengthened online safety culture

---

Future Roadmap

• CLIP-based multimodal filtering (image + text)
• ONNX / TensorRT conversion for faster inference
• User-personalized filtering mode
• Support for mobile browsers

---

Contributors

JEONG AH YOON : Image Interceptor AI Research & Engineering Team

---

License

MIT License © Image Interceptor