Machine Learning in Practice – Detection and Classification (AICLS)

AI - Artificial Intelligence, AI for developers

This course provides a practical framework for building detection and classification systems using machine learning. You will cover the model lifecycle: data preparation and feature engineering, algorithm selection and evaluation metrics (precision/recall, ROC-AUC, F1) and model interpretation.

It focuses on avoiding common mistakes such as data leakage and bad validation, ensures reproducibility and safe data handling, and covers deployment and operations: Docker/REST, monitoring and drift detection, retraining, sample notebooks, pipelines and SOC/MLOps integration.

THIS TRAINING COURSE WILL HELP YOU:

  • Differentiate classic ML and deep learning for detection tasks
  • Select proper metrics and evaluate models (precision/recall, ROC-AUC, F1)
  • Design and build a classification pipeline: features › model › interpretation
  • Deploy models with Docker and CI/CD and monitor model drift

WHO SHOULD ATTEND?

  • Data and ML engineers and analysts
  • SOC and security teams handling detection
  • Backend developers and anyone building automated detection

COURSE LOCATION AND AVAILABLE DATES



This course is not scheduled as a public course.
It is delivered exclusively as customised training. The content, duration, date, and delivery format are tailored to the customer's requirements and can also be provided in English.

Request customised training

Course content:

Hide details
  • Introduction & ML basics
    1. Terminology: supervised/unsupervised, classification vs regression, anomaly detection
    2. ML vs DL: when tree-based models suffice and when CNN/LSTM/Transformers are needed
    3. Data leakage, train/test split, cross-validation — common mistakes and prevention
  • Detection domains (security use-cases)
    1. Phishing & spam: text/URL features, domain reputation
    2. DGA/malicious domains: length, entropy, n-grams, WHOIS/TLS attributes
    3. Log anomalies: outlier detection, DBSCAN, Isolation Forest
    4. Data sources: Alexa/Tranco, PhishTank, internal logs, synthetic data
  • Feature engineering & model selection
    1. Lexical/statistical features for domains/URLs, TF-IDF, hashing trick, embeddings
    2. XGBoost/LightGBM, SVM, RandomForest vs simple CNN/LSTM
    3. scikit-learn pipelines, preventing leakage, reproducibility
    4. Hyperparameter tuning: GridSearchCV vs Optuna
  • Metrics and validation
    1. Confusion matrix, ROC vs PR curve (imbalanced data)
    2. F1, balanced accuracy, MCC
    3. Model interpretation: SHAP/LIME
    4. K-fold CV, stratification, time-based split
  • Deployment and operations
    1. Model export (pickle/joblib vs ONNX)
    2. Docker + REST API (FastAPI/Flask), CI/CD pipeline
    3. Quality monitoring, prediction logging, drift detection, retraining (Airflow/cron), versioning (MLflow/DVC)
  • Hands-on workshop
    1. From CSV to classifier: data prep, feature engineering, model training
    2. Metric visualization (ROC/PR), SHAP interpretation
    3. Mini-deploy (local REST API) + manager-facing report
  • Bonus / extensions
    1. Active learning, semi-supervised approaches
    2. Feature store & model registry (Feast, MLflow)
    3. SIEM/SOC integration, auto-enrichment of alerts
    4. MLOps basics: monitoring, governance, reproducibility
Prerequisites:
Basic Python skills and a basic understanding of machine learning.
Schedule:
2 days (9:00-17:00)

Training and learning environment