Technical Summary: Framework for Discovering GPS Spoofing Attacks in Drone Swarms

1. Problem Statement

Drone swarms, utilized for safety-critical tasks like logistics and search-and-rescue, rely on distributed swarm control algorithms to maintain formation and avoid collisions. While significant research has focused on improving swarm intelligence, the security implications of design choices in these control algorithms remain under-explored.

The paper identifies a specific class of vulnerabilities termed Swarm Propagation Vulnerabilities (SPVs). Unlike traditional attacks that target a single drone or require high-cost external intruders, SPVs exploit the interdependence of swarm members. An attacker can launch a GPS spoofing attack on a single "target drone" with a small, stealthy deviation (e.g., 0–10 meters). This deviation causes the target drone to alter its trajectory to avoid collisions with other members. Consequently, the swarm control algorithm generates incorrect control commands for "victim drones" to maintain formation or cohesion. These incorrect commands cause the victim drones to veer off course, leading to collisions with obstacles or other drones.

Current defenses for single drones often ignore small GPS deviations, assuming they are indistinguishable from standard GPS noise. However, in a swarm, these small deviations propagate through the control logic, causing catastrophic failures that are difficult to attribute to the initial attacker.

2. Methodology

To systematically discover SPVs, the authors propose a fuzzing framework consisting of two tools: SwarmFuzzGraph and SwarmFuzzBinary. Both tools aim to find the optimal "seed" (target-victim drone pair) and "spoofing parameters" (direction, start time, duration) that minimize the distance between a victim drone and an on-path object (obstacle or another drone).

SwarmFuzzGraph

This initial tool employs two heuristics to navigate the large input space:

1. Seed Scheduling via Swarm Vulnerability Graph (SVG): The tool constructs a graph where nodes represent drones and edges represent interactions. Using PageRank centrality, it quantifies the "malicious influence" of each drone. The most influential drone is selected as the target, and the drone closest to an obstacle is selected as the victim. This prioritizes pairs likely to cause collisions.
2. Search-Based Fuzzing via Gradient Descent: Once a seed is selected, the tool treats the collision distance as a convex objective function. It uses gradient descent to efficiently search for the spoofing duration and start time that minimize this distance, aiming to find the specific parameters that trigger a collision.

SwarmFuzzBinary

The authors identified limitations in SwarmFuzzGraph, specifically its lack of generality across different swarm topologies (due to the SVG abstraction) and the sensitivity of gradient descent to parameter tuning (e.g., learning rates). SwarmFuzzBinary addresses these issues:

1. Topology-Agnostic Seed Scheduling: Instead of a graph model, it observes the victim drone's behavior. If a maximum GPS spoofing attack causes the victim drone to pass an obstacle from a different direction compared to the no-attack scenario, the pair is deemed influential. This relies on the property that if a strong attack fails to change the outcome, a weaker one will too.
2. Parameter-Agnostic Search: The tool transforms the objective function into a monotonic function based on the direction of the victim drone relative to the obstacle. It then employs binary search to find the spoofing duration that triggers the collision. This eliminates the need for manual parameter tuning (like learning rates) required by gradient descent.

3. Key Contributions

The paper makes the following contributions:

• Demonstration of SPVs: It is the first work to demonstrate that GPS spoofing on a single drone can indirectly cause collisions in a swarm by exploiting control algorithm logic.
• SwarmFuzzGraph: A fuzzing tool utilizing SVG and gradient-guided optimization to find SPVs.
• SwarmFuzzBinary: An improved fuzzing tool using observation-based seed scheduling and binary search, designed to be robust across varying swarm topologies and requiring no manual parameter tuning.
• Systematic Evaluation: The authors evaluated both tools on two popular swarm control algorithms: Vicsek (focused on formation alignment) and Olfati-Saber (focused on potential field-based obstacle avoidance).

4. Experimental Results

The tools were evaluated in the SwarmLab simulator with swarm sizes of 5, 10, and 15 drones, and GPS spoofing distances of 5m and 10m.

• SwarmFuzzGraph Performance:

  • Achieved an average success rate of 48.8% in finding SPVs in the Vicsek algorithm.
  • The SVG heuristic boosted the success rate by more than 10x compared to random selection.
  • Gradient-guided optimization reduced runtime overhead by 3x compared to random fuzzing.
  • Limitation: It failed to find any SPVs in the Olfati-Saber algorithm due to topology sensitivity and optimization overshooting.

• SwarmFuzzBinary Performance:

  • Achieved an average success rate of 60.2% in Vicsek and 86.5% in Olfati-Saber.
  • Successfully identified SPVs in both algorithms, including collisions between drones (not just with obstacles).
  • Efficiency: Compared to a random fuzzer, SwarmFuzzBinary achieved a 7x higher success rate and 5x lower runtime. The seed scheduling heuristic alone boosted success rates by 5.4x.
  • Observations: Larger swarms and larger spoofing distances correlated with higher success rates. SwarmFuzzBinary found that the Olfati-Saber algorithm is more susceptible to small location disruptions, requiring shorter spoofing durations (avg 1.15s) compared to Vicsek (avg 8s).

5. Significance and Claims

The paper claims to be the first to systematically identify and provide a framework for discovering SPVs in drone swarm control algorithms. Its significance lies in:

• Shifting the Threat Model: Highlighting that small, stealthy physical attacks (GPS spoofing) can be more debilitating than jamming or logic flaws because they exploit the swarm's own coordination mechanisms.
• Proactive Defense: Providing a tool for designers to evaluate swarm resilience before deployment. If vulnerabilities are found, control parameters can be tuned to mitigate them.
• Algorithmic Insights: The evaluation reveals that algorithms prioritizing formation maintenance (Vicsek) or potential field methods (Olfati-Saber) have distinct failure modes under sensor attacks, such as "trap scenarios" where drones get stuck, or "trace intersections" causing collisions.
• Design Implications: The authors suggest that future swarm designs must account for sensor noise and small spoofing deviations, potentially requiring stricter protection techniques for large swarms with low distances to obstacles (VDO).

The paper concludes that while SwarmFuzzGraph was effective for specific topologies, SwarmFuzzBinary offers a more robust, generalizable approach to finding these vulnerabilities without requiring manual tuning or specific knowledge of the swarm topology.