📊 Data & Models

This section covers how behavioral data is used to train machine learning models in bezpieczenstwoPwr, as well as how those models are integrated and run in the browser.

📁 Data Source

The machine learning models are trained using anonymized user behavior data, primarily from:

• Kaggle datasets related to mouse movement, bot detection, or human-computer interaction
• Custom-collected data from real users and simulation tools (e.g., puppeteer)

Typical features include:

• Mouse speed and acceleration
• Click patterns and hesitation times
• Pointer trajectory shapes
• Keystroke intervals and input dynamics

🧠 All data is anonymized before any storage or model training. No personally identifiable information is collected.

🛠 Data Preprocessing

Before feeding data into the model, we perform several preprocessing steps:

• Normalization of coordinates and timing
• Sequence padding or trimming to ensure consistent input shape
• Feature extraction such as velocity, direction changes, or dwell time
• Noise filtering to remove accidental or edge-case inputs

You can find preprocessing code inside the training pipeline scripts (not yet published, ask us if you want early access).

🧠 Model Architecture

Models are designed for lightweight, in-browser inference using TensorFlow.js.

We experimented with:

• LSTM / GRU networks for time-sequence analysis
• 1D CNNs for spatial-temporal patterns
• Ensemble methods trained offline and distilled into TensorFlow.js format

The final model is selected based on size, accuracy, and inference speed in the browser.

⚠️ Models are kept under public/models/ for now. Expect a proper versioning system and CDN-hosted models in future updates.

📦 Loading Models

The bbotd package automatically loads a model from a given URL.