kodzuken

AI-Assisted Network Intrusion Detection (MESRS)

Machine Learning / Network Security

An AI-assisted intrusion detection system for the information security team of Algeria's Ministry of Higher Education (MESRS). It reads how traffic behaves rather than what it contains, so it works on encrypted connections, never sits in the production path, and raises only as many alerts as the team can actually review.

AI-Assisted Network Intrusion Detection (MESRS) - Machine Learning / Network Security cover image

Project Overview

The ministry runs online platforms used by the whole Algerian higher-education sector, and like any public service they receive automated scans, floods and intrusion attempts. The usual protections judge one connection or one server at a time, but some attacks only show across many connections: a machine probing hundreds of addresses, a server being slowly worn down, a compromised host quietly calling home every few minutes. Most of that traffic is also encrypted, so its content cannot be read.

The brief was open: find where AI would help the security team most. I chose network intrusion detection, wrote the proposal, and had its objectives approved before building anything. The system reads a copy of the traffic, so production is never touched, and it decrypts nothing. It works only from what encryption leaves visible: who talks to whom, when, how much, for how long, and how each connection ends.

It looks at the traffic from three points of view, each source machine, each server, and each pair of machines over time, and then asks the question that mattered most: is this unusual for this particular host? Related signals are grouped into incidents that carry a severity, the reasons behind it and the evidence, and the alert threshold is set from the number of incidents the analysts can actually review, not from a score that only looks good on paper.

Key Features

  • Passive and detection-only: it reads a mirrored copy of the traffic, so it can never slow down or block a legitimate user
  • Works on encrypted traffic without decrypting it, using only connection metadata: endpoints, timing, volumes, how a connection ends and the parameters of its handshake
  • Three complementary detectors: one follows each source machine (scanning, spreading), one gathers every client of a server (floods, slow denial of service, password guessing), one follows a pair of machines over time (a compromised host calling its controller)
  • A context layer that judges each host against its own normal behavior, so a machine whose job legitimately looks like scanning, such as a monitoring server, does not bury the team in alerts
  • Incidents instead of raw alerts, each with a severity, its reasons and its evidence, delivered into the monitoring tool the analysts already use
  • An alert budget: the threshold is calibrated to how many incidents the security team can review, so every alert is worth opening
  • Privacy by design: a data-collection kit pseudonymizes addresses with a key that never leaves the ministry, so no model ever sees a real address
  • Built and validated on public research datasets first, so the ministry's own traffic only had to be plugged in

Technologies Used

  • Python
  • pandas
  • Machine Learning
  • Anomaly Detection
  • Network Traffic Analysis
  • Network Security

Project Gallery

More Projects

Back to portfolio