Skein

⏺ Skein is a self healing drone swarm mesh that uses real ML to detect cyberattacks and jamming in real time, rerouting around the damage before the complete failure. Detects real internet traffic.

Comment

Inspiration

Modern defense relies heavily on drone swarms, which are highly vulnerable to radio jamming and cyberattacks. Skein is built to counter this threat: a self-healing swarm that detects interference using real machine learning and autonomously reroutes around damaged nodes.

What It Does

  • Resilient Mesh: 13 drones maintain a dense network where each node connects to at least 4 neighbors.
  • Live Intrusion Detection: A classic ML model analyzes traffic patterns to detect jamming or hacks.
  • Autonomous Rerouting: The swarm uses shortest-path routing to instantly heal when a node drops.
  • Interactive Demo: Attackers can launch strikes from a secondary physical laptop over a hotspot, watching the swarm dynamically adapt when the laptop lid is closed.

How I Built It

  • Frontend: Next.js, React, TypeScript, and Framer Motion for live animations.
  • Backend & Routing: Python, FastAPI, WebSockets for state streaming, and NetworkX for graph routing.
  • ML Detector: An XGBoost model trained on the CIC-IDS-2017 dataset, analyzing 10 flow features to classify traffic (Benign, DoS, Port Scan, Brute Force).

Challenges & Solutions

  • Real-State Simulation: Avoided faking data; jamming simulates real link degradation carrying actual attack traffic.
  • Multi-Device Sync: Overcame hotspot latency and heartbeat timing issues to seamlessly integrate the second physical laptop.
  • Mesh Tuning: Iterated on graph density to ensure the network remains stable even after losing multiple nodes simultaneously.

Accomplishments

  • Achieved 99% accuracy on unseen evaluation data, backed by live confusion matrices and feature importance metrics.
  • Created a fully functional, two-laptop hardware demo where physical disconnection triggers immediate, authentic network healing.

Lessons Learned

  • Tree Models > Deep Learning: XGBoost outperformed neural networks on tabular network data while providing crucial feature explainability.
  • Authenticity Wins: Keeping the data generation and simulation honest provided massive credibility during live demonstrations.

What's Next

  • Anomaly Detection: Transition to unsupervised learning to flag novel, zero-day attacks.
  • Live Packet Processing: Drive the detector using real-time packet capture rather than replayed dataset rows.
  • Scale: Expand past 13 nodes and harden the routing algorithm against multi-vector concurrent attacks.

Built With

Share this project:

Updates