DUCK: A Drone-Urban Cyber-Defense Framework Based on Pareto-Optimal Deontic Logic Agents

Tonmoay Deb, Jürgen Dix, Mingi Jeong, Cristian Molinaro, Andrea Pugliese, Alberto Quattrini Li, Eugene Santos Jr., V. S. Subrahmanian, Shanchieh Yang, Youzhi ZhangAAAI 2023 · DOI

TL;DR

This paper introduces DUCK, a novel multi-agent testbed for simulating and defending against drone-based urban terrorist attacks. DUCK leverages Pareto-optimal deontic logic agents to model complex defensive strategies involving CCTVs, blue drones, and cyber-attacks, enabling security agencies to test and optimize their responses.

What Problem Does This Paper Address?

The increasing threat of drone-based attacks in urban environments poses a significant challenge for security agencies. These attacks can be carried out by diverse actors with varying capabilities, making it difficult to develop effective, comprehensive defense strategies. Existing methods often struggle to simulate the complex interactions between multiple defensive assets (like drones and surveillance systems) and offensive drone swarms, especially when considering ethical and operational constraints on defensive actions. The authors aim to fill this gap by providing a flexible and realistic simulation environment that allows for the exploration of various defense scenarios and the optimization of defensive resource allocation.

Methods

DUCK is built upon a deontic logic theory, which allows agents to reason about what they ought to do, may do, or are forbidden from doing. This is crucial for modeling the constraints and permissions of defensive agents. The core of DUCK's approach lies in its Pareto-optimal deontic logic agents. At each time step, each agent computes a set of actions that are optimal across multiple objectives (e.g., minimizing damage to protected areas, conserving battery) and are consistent with their deontic logic "agent program." This means agents don't just pick a single best action, but a set of equally good, non-dominated options.

The testbed itself is implemented as a graph where nodes represent city regions and edges represent adjacency. Autonomous agents, including Blue and Red HQs, Blue and Red drones, and CCTV agents, operate within this environment. DUCK uses Microsoft's AirSim simulator for realistic 3D visualization and physics, extended with features like drone-drone attacks, destruction, hacking, and battery/payload constraints. The system is orchestrated using the Robot Operating System (ROS) for communication, decision-making, control, and visualization.

Datasets

Not discussed in the paper. The DUCK testbed is a simulation environment, and the paper does not mention training or evaluating on specific real-world datasets.

Key Insights and Ideas

  • Deontic Logic for Realistic Defense: Using deontic logic to define agent behavior (permissions, prohibitions, obligations) is a powerful way to model the nuanced rules of engagement for defensive systems, going beyond simple objective optimization.
  • Pareto-Optimality for Multi-Objective Decision Making: Agents don't just aim for one goal; they consider multiple objectives simultaneously. Computing Pareto-optimal sets of actions ensures that defenders explore a range of "best possible" strategies, rather than a single, potentially suboptimal, choice.
  • Integrated Cyber-Physical Defense Simulation: DUCK's ability to combine physical drone combat with cyber-attacks (hacking) and surveillance (CCTVs) provides a holistic view of urban defense, reflecting real-world complexities.
  • Agent-Based Simulation for "What-If" Scenarios: The testbed allows security agencies to ask critical questions like "how many blue drones are needed to save X% of New York City?" or "what is the impact of investing Y dollars in defense?" This is invaluable for strategic planning and resource allocation.

Key Results

The paper focuses on the framework and testbed itself rather than presenting quantitative results from extensive simulations. The key "result" is the successful development and demonstration of the DUCK testbed, which is described as having over 10,100 lines of code in C++ and Python. The demonstration showcases the visualization of a simulated urban environment with blue and red drones, CCTVs, and HQ views, illustrating the system's capabilities in representing complex scenarios.

Limitations and Future Work

  • Limitations: The paper doesn't explicitly detail limitations of the current DUCK implementation, but the complexity of simulating real-world urban environments and the stochastic nature of agent actions imply that perfect prediction is impossible. The focus on agent programs and Pareto-optimality is a significant computational undertaking.
  • Future Work: The authors plan to extend DUCK by incorporating game-theoretic and reinforcement learning models, further developing hacking capabilities, and exploring pursuit-evasion and human-agent interaction techniques. This suggests a roadmap for making the simulation even more sophisticated and realistic.

Summary updated

BibTeX
@inproceedings{deb2023duck,
  author = {Tonmoay Deb and Jürgen Dix and Mingi Jeong and Cristian Molinaro and Andrea Pugliese and Alberto Quattrini Li and Eugene Santos Jr. and V. S. Subrahmanian and Shanchieh Yang and Youzhi Zhang},
  title = {{DUCK: A Drone-Urban Cyber-Defense Framework Based on Pareto-Optimal Deontic Logic Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.27067},
}
Figure