Kotlin TensorFlow Lite Platform
YBrowser is an experimental Android web browser that ditches traditional static blocklists (like EasyList) in favor of a local, on-device Artificial Intelligence.
Instead of searching through a massive text file of millions of blocked URLs, YBrowser uses a Multi-Modal Neural Network (TFLite) to mathematically predict if a network request is an ad or a tracker in real-time, executing in under 2 milliseconds per request.
 |
 |
|---|---|
| OFF | ON |
Traditional ad blockers rely on static lists (e.g., ||ads.example.com^). This approach has three major flaws:
- Memory Bloat: Loading hundreds of thousands of rules into Android's RAM is highly inefficient.
- Cat-and-Mouse Game: If an ad network changes their domain to new-ads.com, the static list breaks until a human manually updates it.
- Zero-Day Trackers: Trackers using Domain Generation Algorithms (DGAs) bypass static lists entirely.
The Solution: YBrowser extracts 27 real-time features from every HTTP request made by the WebView and feeds them into a quantized AI model. The AI recognizes the behavior and mathematical structure of an ad, allowing it to block trackers it has never even seen before.
The AI Architecture & Feature Engineering
The underlying AI model analyzes each network request across three distinct categories:
- Lexical & Mathematical Features
Trackers often use randomly generated domains or incredibly long URL paths. The AI extracts:
- Shannon Entropy: Calculates the randomness of the URL characters. High entropy often indicates a DGA tracker.
- Path Depth: Counts the depth of the directory structure (/wp-content/uploads/ads/).
- Query Parameters: Analyzes the size and structure of tracking tags appended to URLs.
- Contextual & Network Features
- First-Party vs. Third-Party: By comparing the requested URL to the documentUrl (the page the user is currently looking at), the AI heavily penalizes third-party background requests.
- Resource Type: Evaluates if the WebView is requesting a script, image, stylesheet, or xmlhttprequest. Ads are almost entirely skewed towards scripts and images.
- Heuristic & Regex Features
- Dimensionality Checks: Looks for standard IAB ad unit sizes inside the URL (e.g., 300x250, 728x90).
- Ad-Tech Keywords: Scans for subtle inclusions of words like telemetry, beacon, or pixel at the boundaries of the URL path.
Building an AI into a mobile browser requires extreme optimization to prevent the UI from freezing.
- Custom Standard Scaler: Neural networks require normalized inputs (between 0 and 1). Instead of using heavy Python libraries, YBrowser hardcodes the means and scales directly into Kotlin, performing (X - Mean) / Scale in microseconds.
- Thread-Safe Inference: Android WebView triggers shouldInterceptRequest concurrently across dozens of background threads. The inference engine is wrapped in a synchronized(aiLock) block, preventing the XNNPack memory allocator from crashing under concurrent loads.
- Main Thread UI Updates: Blocked ad counters are safely dispatched to the UI using lifecycleScope.launch { withContext(Dispatchers.Main) { ... } }.
- Zero-Latency Dropping: When the AI flags a request (output > 0.5f), the browser intercepts it by returning an empty WebResourceResponse, saving the user's bandwidth instantly.
- Clone the repository: git clone https://github.com/raslenabb12/YBrowser-AI.git
- Open the project in Android Studio.
- Ensure the ad_blocker_quant.tflite model is present in the app/src/main/assets/ directory.
- Build and run on your Android device or emulator.
As an experimental ML model, you may occasionally encounter False Positives. Adjusting the probability threshold (output > 0.5f) can tune the aggressiveness of the blocker.
This project is licensed under the MIT License - see the LICENSE file for details.