Population Anomaly Detection with PyTorch
summary
Implements an adversarial autoencoder that forces the latent space to match a unit Gaussian so high‑dimensional distributions become axis‑independent and divergences are easy to compute. Uses PyTorch to train and decode, applies to DNS exfiltration and website visitor logs, and produces per‑transaction risk scores and 168‑hour heatmaps for insight. ROC curves show reliable detection down to about 0.1% injection and clear separation at 10% and 1%, though the method currently lacks feature‑level explanations.
AI-written summary · about
chapters (11)
- 0:00Introduction to anomalies
- 2:14Defining soft population anomalies
- 7:55Curse of dimensionality
- 10:03Autoencoders for anomaly detection
- 12:38Adversarial encoders explained
- 16:55Implementing the solution
- 17:16PyTorch overview
- 21:36User activity analysis
- 24:04Data exfiltration detection
- 25:51Risk scoring and results
- 26:53Future directions
related talks
45mAnomaly detection with TensorFlow
48mHunting Stealth Adversaries with Graphs & AI
28mAnomaly Detection using Autoencoders
26mAnomaly Detection System for the CERN Cloud
31mIndividual Gaps in Prediction Generalization
01h 01mD1T2 - Applying Machine Learning to User Behavior Anomaly Analysis
25mMonitoring AI with AI
35mAnomaly Detection in Smart Buildings using Federated Learning