Robust and Efficient AI-Based Attack Recovery in Autonomous Drones

TL;DR

The paper addresses the need for robust, real-time attack recovery in autonomous drones operating under dynamic adversarial conditions. It proposes a hierarchical recovery framework that fuses a lower-level simplex-based control loop with a high-level GenAI planner that outputs target sets $T \in \mathcal{T}$ and parameters $\theta \in \Theta$, enabling adaptive, safe recovery actions. A safety/feasibility verifier and a Multi-modal LLM assist in selecting safe target zones, with robustness enhanced through randomized smoothing to withstand adversarial inputs. To enable edge deployment, the approach leverages model distillation, linear-time architectures, and post-training quantization, aiming to maintain real-time performance while mitigating GPS spoofing and sensor attacks.

Abstract

We introduce an autonomous attack recovery architecture to add common sense reasoning to plan a recovery action after an attack is detected. We outline use-cases of our architecture using drones, and then discuss how to implement this architecture efficiently and securely in edge devices.

Figures (5)

• Figure 1: A drone receives false GNSS information, forcing it to lower it's altitude. OPR detects this attacks and returns the drone to a safe altitude.
• Figure 2: Our algorithm (OPR-OL) returns a drone to a safe height (green area) faster and more accurately than previous work. In addition, if we can filter out the malicious sensor and take the input from the remaining sensors, we obtain a Partially Closed Loop (OPR-PCL) algorithm that outperforms slightly our open loop model.
• Figure 3: Success rate and average distance to the target set center with increasing noise for the drone.
• Figure 4: AI-Based Recovery.
• Figure 5: Multi-modal LLM (MLLM) evaluates the risks and ranks possible safe landing locations.