Machine Learning for Fraud Detection in Streaming Services

Soheil Esmaeilzadeh, Negin Salajegheh, Amir Ziai, Jeff Boote21 Sep 2022

Problem

A streaming service limits active devices, concurrent streams and downloaded titles. Millions of accounts on many device types give a large attack surface: content fraud, service fraud and account fraud. Netflix security teams needed to catch these at scale and in near real time. Rule-only detection needs constant expert upkeep and does not scale. Model-based detection needs labels, and Netflix had neither a trained model nor labeled fraud examples.

Example from the article: a device that tries many parameter combinations leaves a long trail of streaming errors, which benign devices do not.

Solution in brief

Security experts write heuristics that tag suspicious accounts automatically. The tags become labels for daily account-level features (counts and shares of devices, DRM types, formats, licenses). Models trained on these labels, a deep autoencoder trained on benign accounts only and gradient-boosted classifiers, are meant to learn patterns beyond the heuristics and to correct their false positives.

Questions this case answers
  • How to build a fraud detector for account-level behavior when no labeled fraud exists and only expert rules are available
  • How to find which behavioral signals separate abusive accounts from normal ones when abuse has several overlapping types
  • How to detect account, service and content fraud in a streaming service when there are no labeled fraud examples
  • How to turn security experts' rules into training labels for a fraud model

Tools and methods

  • Heuristic labeling — expert rules tag anomalous accounts; a vetted trusted group is the benign class
  • Deep autoencoder — semi-supervised model trained on benign accounts only, flags high reconstruction error
  • One-Class SVM, Isolation Forest, Elliptic Envelope, Local Outlier Factor — one-class baselines
  • SMOTE — oversampling of the minority classes
  • XGBoost and other classifiers — binary and multi-class multi-label fraud detection
Continue reading — it's free

Google or a one-time email code. No password.