Quadcopter Cybersecurity and Anomaly Detection

Overview

This project is a complete machine learning solution for anomaly detection in quadcopters, addressing cybersecurity threats. It covers the entire machine learning pipeline, from data generation to novel model development, ensuring enhanced security against cyber-attacks.

Project Goals

• Provide higher security for quadcopters and drones against malicious cyber-attacks.
• Generate and collect data using MATLAB and Simulink.
• Develop novel deep learning models for cyber-attack detection, identification, and isolation.
• Extend solutions to multi-quadcopter networks for robust anomaly detection.

Data Generation

1. Quadcopter Modeling & Control

• Quadcopter modeled from mathematical formulas using Simulink's State-Space block.
• Implemented both linear and non-linear models.
• Open-loop system tested before designing closed-loop controllers.
• PID controllers (single & cascaded) used for 3D control over x, y, and z axes.
• Non-linear quadcopter modeled using fundamental Simulink blocks (e.g., sin, cos functions).
• Same controller applied to both linear & non-linear models.
• Successfully implemented a fully maneuverable, non-linear quadcopter with local control.

2. Cybersecurity (Cyber-Attack Simulation)

• Implemented three cyber-attacks:
  • Denial of Service (DoS)
  • False Data Injection (FDI)
  • Replay Attack
• Cyber-attacks implemented using MATLAB-Simulink combinations.
• Utilized Zero-Order Hold, Memory, and Delay blocks to introduce attacks.
• Attacks applied to both sensors and actuators.

3. Scenario-Based Data Collection

• Defined movement scenarios for quadcopter flight paths.
• Monte Carlo approach used to generate 50 randomized scenarios.
• Parameters randomized include attack timing, movement paths, etc.
• Sensor and actuator data collected from Simulink to MATLAB, stored as CSV files.

Machine Learning for Cyber-Attack Detection

1. Base Model Selection

• RNN-based models selected due to the time-series nature of the data.
• LSTM (Long Short-Term Memory) chosen as the best RNN-based model.
• Basic LSTM model implemented in Python (Google Colab) for initial cyber-attack detection.

2. Novel Model Development

• Multi-Output LSTM (MO-LSTM) introduced with a shared LSTM backbone and 3 output heads for:
  • Cyber-attack detection
  • Attack type identification
  • Affected component isolation
• Models developed using TensorFlow and Keras.
• MO-LSTM model successfully implemented and tested.

Extension to Multi-Quadcopter Networks

1. Dataset Generation for Networks

• Number of quadcopters increased from 1 to 5 in Simulink & MATLAB.
• Adjusted Simulink models and MATLAB scripts accordingly.
• 50 new Monte Carlo scenarios generated for multi-quadcopter simulations.
• Sensor & actuator data transferred to MATLAB and converted into CSV files with automatic attack labels.

2. MI-MO LSTM Model (Multi-Input, Multi-Output LSTM)

• Previous MO-LSTM model worked well for single quadcopter with 10 features.
• To handle multiple quadcopters, a new adaptable model was designed:
  • Multiple input heads for each quadcopter's sensor & actuator data.
  • Shared LSTM backbone to extract common patterns.
  • Three output heads for detection, identification, and isolation.
• Tested with 2 to 5 quadcopters – results showed strong detection performance.
• Model named MI-MO LSTM (Multi-Input, Multi-Output LSTM).

Results & Future Work

• High detection accuracy achieved for both single and multi-quadcopter scenarios.
• Model successfully detects, identifies, and isolates cyber-attacks.
• Future work includes:
  • Deploying models on real-world UAVs.
  • Expanding dataset with additional attack types.
  • Optimizing real-time processing capabilities.