mansi14883md/AI-Keystroke-Security-System

1import json2import os3 4import joblib5import pandas as pd6from sklearn.ensemble import RandomForestClassifier7from sklearn.ensemble import ExtraTreesClassifier8from sklearn.metrics import accuracy_score, classification_report, f1_score, precision_score, recall_score9from sklearn.model_selection import StratifiedKFold, cross_val_score10from sklearn.model_selection import train_test_split11from sklearn.pipeline import Pipeline12from sklearn.preprocessing import StandardScaler13from sklearn.svm import SVC14 15from backend_utils import engineered_feature_columns, transform_dataframe16 17 18DATA_FOLDER = "data"19MODEL_FOLDER = "models"20GENUINE_FILE = os.path.join(DATA_FOLDER, "genuine.csv")21IMPOSTOR_FILE = os.path.join(DATA_FOLDER, "impostor.csv")22MODEL_FILE = os.path.join(MODEL_FOLDER, "model.pkl")23METRICS_FILE = os.path.join(MODEL_FOLDER, "metrics.json")24 25 26def load_dataset() -> pd.DataFrame:27    if not os.path.exists(GENUINE_FILE):28        raise FileNotFoundError("genuine.csv not found. Run capture_data.py first.")29    if not os.path.exists(IMPOSTOR_FILE):30        raise FileNotFoundError("impostor.csv not found. Run capture_data.py first.")31 32    genuine_df = pd.read_csv(GENUINE_FILE)33    impostor_df = pd.read_csv(IMPOSTOR_FILE)34    return pd.concat([genuine_df, impostor_df], ignore_index=True)35 36 37def build_candidate_models() -> dict[str, object]:38    return {39        "Random Forest": RandomForestClassifier(n_estimators=200, random_state=42),40        "Extra Trees": ExtraTreesClassifier(n_estimators=300, random_state=42),41        "SVM": Pipeline(42            [43                ("scaler", StandardScaler()),44                ("classifier", SVC(kernel="rbf", probability=True, random_state=42)),45            ]46        ),47    }48 49 50def main() -> None:51    os.makedirs(MODEL_FOLDER, exist_ok=True)52 53    df = load_dataset()54    if len(df) < 10:55        raise ValueError("Dataset is too small. Record more samples before training.")56 57    feature_df = transform_dataframe(df)58    X = feature_df.drop(columns=["label"])59    y = feature_df["label"]60 61    X_train, X_test, y_train, y_test = train_test_split(62        X,63        y,64        test_size=0.3,65        random_state=42,66        stratify=y,67    )68 69    candidate_models = build_candidate_models()70    trained_models = {}71    model_scores = {}72    cv_scores = {}73    cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)74 75    for model_name, model in candidate_models.items():76        scores = cross_val_score(model, X, y, cv=cv, scoring="accuracy")77        cv_scores[model_name] = float(scores.mean())78        model.fit(X_train, y_train)79        predictions = model.predict(X_test)80        model_scores[model_name] = accuracy_score(y_test, predictions)81        trained_models[model_name] = model82 83    best_model_name = max(cv_scores, key=cv_scores.get)84    model = trained_models[best_model_name]85 86    # Fit the selected model again on the full dataset for stronger final deployment.87    model.fit(X, y)88 89    evaluation_model = trained_models[best_model_name]90    predictions = evaluation_model.predict(X_test)91    accuracy = accuracy_score(y_test, predictions)92    precision = precision_score(y_test, predictions, zero_division=0)93    recall = recall_score(y_test, predictions, zero_division=0)94    f1 = f1_score(y_test, predictions, zero_division=0)95 96    metrics = {97        "accuracy": round(float(accuracy), 4),98        "precision": round(float(precision), 4),99        "recall": round(float(recall), 4),100        "f1_score": round(float(f1), 4),101        "train_samples": int(len(X_train)),102        "test_samples": int(len(X_test)),103        "best_model": best_model_name,104        "model_comparison": {name: round(float(score), 4) for name, score in model_scores.items()},105        "cross_validation_accuracy": {name: round(float(score), 4) for name, score in cv_scores.items()},106        "feature_count": len(engineered_feature_columns()),107        "decision_threshold": 0.5,108    }109 110    model_bundle = {111    "model": model,112    "feature_columns": engineered_feature_columns(),113    "best_model": best_model_name,114    "decision_threshold": 0.5,115    "target_text": "secure123",116    }117 118 119    joblib.dump(model_bundle, MODEL_FILE)120    with open(METRICS_FILE, "w", encoding="utf-8") as file:121        json.dump(metrics, file, indent=4)122 123    print("Model training completed successfully.")124    print(f"Model saved as: {MODEL_FILE}")125    print(f"Metrics saved as: {METRICS_FILE}")126    print(f"Best model selected: {best_model_name}")127    print(f"Total engineered features: {len(engineered_feature_columns())}")128    print("\nModel comparison:")129    for model_name, score in model_scores.items():130        print(f"- {model_name}: {score * 100:.2f}%")131    print("\nCross-validation accuracy:")132    for model_name, score in cv_scores.items():133        print(f"- {model_name}: {score * 100:.2f}%")134    print("\nClassification Report:")135    print(classification_report(y_test, predictions, target_names=["Impostor", "Genuine"], zero_division=0))136    print(f"Accuracy: {metrics['accuracy'] * 100:.2f}%")137 138 139if __name__ == "__main__":140    main()141