| --- |
| license: mit |
| library_name: pytorch |
| tags: |
| - keystroke-dynamics |
| - lstm |
| - cybersecurity |
| - hid |
| --- |
| |
| # LSTM-16win-Keystrokes |
|
|
| ## Summary |
| This repository contains a PyTorch **LSTM** classifier for **Human vs HID** keystroke control detection using **windowed** timing features. |
| The label for each window is the **last keystroke** label in that window. |
|
|
| ## Training setup (as implemented) |
| - **Window size:** 16 |
| - **Stride:** 1 |
| - **Label mapping:** {"human": 0, "hid": 1} |
| - **Window label:** last-keystroke in the window |
| - **Features (F=10):** |
| - `duration` |
| - `DD_time` |
| - `DU_time` |
| - `UD_time` |
| - `UU_time` |
| - `run_avg_duration` |
| - `run_avg_DD` |
| - `run_avg_DU` |
| - `run_avg_UD` |
| - `run_avg_UU` |
|
|
| ### Preprocessing |
| - Windows built with NumPy `sliding_window_view`. |
| - Standardization via `StandardScaler` fitted on **training windows only**, across all timesteps and samples. |
|
|
| ### Model |
| - `torch.nn.LSTM` (unidirectional, batch_first) |
| - Hidden size: 64 |
| - Num layers: 1 |
| - Dropout: 0.0 |
| - Head: Linear(hidden_size → 2) |
|
|
| ### Optimization |
| - Optimizer: Adam |
| - LR: 0.001 |
| - Batch size: 256 |
| - Epochs: 30 |
| - Seed: 42 |
|
|
| ## Files |
| - `model.safetensors`: model weights |
| - `config.json`: architecture + feature metadata |
| - `scaler.joblib`: fitted StandardScaler |
| - `metrics.json`: classification report + confusion matrix |
| - `inference.py`: minimal loading + prediction example |
|
|
| ## Usage (minimal) |
| ```python |
| from inference import load_model_and_scaler, predict_df |
| |
| model, scaler, cfg = load_model_and_scaler("NourFakih/LSTM-16win-Keystrokes") |
| y_pred = predict_df(df, model, scaler, cfg) # df must contain cfg["feature_cols"] |
| ``` |
|
|
