A large language model learns to control a drone swarm without revealing sensitive data

A new AI development brings the capabilities of large language models to the use of drone swarms without the need to disclose surveillance data in plain text. The framework, named PrivLLMSwarm, combines swarm control of drones and privacy-protecting methods for surveillance tasks in Internet of Things (IoT) environments.

The goal is to utilize large language models that can both make inferences and coordinate actions in natural language. Until now, the problem has been that the information collected by drones – such as camera images or identification data – had to be fed to the AI in plain text, making it susceptible to misuse.

PrivLLMSwarm solves this by using secure multi-party computation. In this process, multiple parties can jointly run an AI model so that no one sees the entire input data or intermediate results, but the final result – such as a command on where the drones should move – is still accessible.

The developers have modified the structure of the transformer-based language model to better suit encrypted computation. Specifically, non-linear activation functions have been replaced with efficient approximations to keep computation light on platforms flying with limited resources.

At the core of the system is a fine-tuned GPT-based command generator that produces operational instructions for the drone swarm. It has been further developed using reinforcement learning to ensure that the commands better support safe and efficient collaboration.

According to the research, the approach makes AI inference in encrypted form more practical for aircraft that need to operate both securely and with limited computational resources.

Source: PrivLLMSwarm: Privacy-Preserving LLM-Driven UAV Swarms for Secure IoT Surveillance, ArXiv (AI).