Go back to Contents
#Project

Two-Tier ML-Based NIDS for Industrial IoT

#AI Safety #Adversarial ML #Federated Learning #IIoT Security

My Bachelor’s thesis project, built as an Undergraduate Research Fellow at the ICOM Research Group, University of Oviedo. The system is a two-tier network intrusion detection pipeline for Industrial IoT:

  • A federated binary classifier running at the edge (tested on a Raspberry Pi 4, 8GB) flags anomalous traffic locally, without any plant site having to expose raw process data to a central server.
  • A centralized multiclass classifier takes the flagged traffic and identifies the specific attack category, aggregating signal across sites while preserving per-site privacy.

The edge classifier reaches 98.6% precision and 95.0% recall on commodity hardware, showing that privacy-preserving, edge-deployable intrusion detection is practical for industrial environments — not just a datacenter-scale technique.

The project is also where my interest in AI safety sharpened: a false negative here is a missed intrusion, and a false positive is an operator learning to ignore the alarm. Getting the robustness/calibration trade-off right under adversarial and distribution-shifted traffic is a small, concrete version of the same problem alignment and interpretability work is trying to solve at a larger scale.