Back to projects

Mirai Lab research

Packet-Level Behavioural Analysis for Early Identification of Malicious Encrypted Connections

This work presents a lightweight method that detects malicious encrypted network traffic from the first few packets of a connection. It uses only packet timing and size patterns, with no decryption, and reaches high accuracy with low false alarm rates on three types of attack traffic.

Running · From February 8, 2026
  • Encrypted traffic
  • malicious traffic detection
  • early detection
  • time–frequency features
  • lightweight model
  • network security

This work presents a lightweight method that detects malicious encrypted network traffic from the first few packets of a connection. It uses only packet timing and size patterns, with no decryption, and reaches high accuracy with low false alarm rates on three types of attack traffic.

Research approach

  1. Step 1

    Define

    Frame the problem, requirements, and evaluation criteria for the research initiative.

  2. Step 2

    Build

    Develop models, systems, and experiments aligned with the project objectives.

  3. Step 3

    Validate

    Evaluate outcomes through pilots, benchmarks, or field studies before wider deployment.

Overview

Encryption protects user privacy, but it also lets malicious traffic hide from traditional inspection tools. This work proposes a lightweight, decryption-free method for detecting malicious encrypted traffic at the start of a connection.

The method observes only the first 5 to 25 packets of each flow. From these, it extracts simple time-based and frequency-based features of packet sizes, arrival times and payload randomness. A compact neural network with about 3,500 parameters then classifies the flow. Because the model is small and makes its decision early, it suits real-time use on live networks. The method was evaluated on three public datasets: DNS data exfiltration: F1 score of 0.998. Botnet communication: F1 score of 0.92. DNS-over-HTTPS tunnelling: F1 score of 0.997, with a 0.3% false positive rate. In a stricter test, whole capture files were held out from training. The method still achieved an F1 score of 0.989 with a false positive rate under 2%. These results suggest that a connection’s first few packets carry enough information to identify malicious encrypted traffic reliably and efficiently. Future work will focus on collecting a custom dataset from a real network environment and implementing the proposed method on it to validate its practical effectiveness..

Team

Project lead: Sanjana Islam

Publications and datasets

Research papers, models, and open datasets will be shared here as they are released.

Project updates

Milestones, fieldwork notes, and lab announcements will appear on our notices page.

View notices →