LLM-Enabled In-Context Learning for Data Collection Scheduling in UAV-assisted Sensor Networks

TL;DR

This work tackles data collection scheduling in UAV-assisted sensor networks for emergencies, where traditional DRL methods struggle with efficiency and sim-to-real gaps. It introduces ICLDC, an edge-enabled framework that uses in-context learning to translate logged environmental data into natural-language task descriptions and executable data-collection schedules, safeguarded by a safety verifier and enhanced by a perplexity-based exploration to detect jailbreaking attempts. The approach demonstrates reduced packet loss compared with baselines and provides a clear path to real-time, adaptable decision-making in SAR-like scenarios, while also highlighting vulnerabilities of LLMs to adversarial prompts and the need for robust defenses. Overall, ICLDC represents a significant step toward practical, safe, and adaptive UAV scheduling in resource-constrained, mission-critical networks, with future work aimed at real-world deployment and multimodal prompt integration.

Abstract

Unmanned Aerial Vehicles (UAVs) are increasingly being utilized in various private and commercial applications, e.g., traffic control, parcel delivery, and Search and Rescue (SAR) missions. Machine Learning (ML) methods used in UAV-Assisted Sensor Networks (UASNETs) and, especially, in Deep Reinforcement Learning (DRL) face challenges such as complex and lengthy model training, gaps between simulation and reality, and low sampling efficiency, which conflict with the urgency of emergencies, such as SAR missions. In this paper, an In-Context Learning (ICL)-Data Collection Scheduling (ICLDC) system is proposed as an alternative to DRL in emergencies. The UAV collects sensory data and transmits it to a Large Language Model (LLM), which creates a task description in natural language. From this description, the UAV receives a data collection schedule that must be executed. A verifier ensures safe UAV operations by evaluating the schedules generated by the LLM and overriding unsafe schedules based on predefined rules. The system continuously adapts by incorporating feedback into the task descriptions and using this for future decisions. This method is tested against jailbreaking attacks, where the task description is manipulated to undermine network performance, highlighting the vulnerability of LLMs to such attacks. The proposed ICLDC significantly reduces cumulative packet loss compared to both the DQN and Maximum Channel Gain baselines. ICLDC presents a promising direction for intelligent scheduling and control in UASNETs.

Figures (5)

• Figure 1: Overall design of the proposed ICLDC. The UASNETs interact with an edge-hosted LLM through structured prompts. The LLM receives logged environmental data (e.g., queue length, channel conditions, and battery level) and generates optimized data collection schedules. A feedback loop records system performance and guides continuous adaptation to minimize packet loss.
• Figure 2: The proposed ICLDC uses a data communication protocol where each communication frame queries the LLM to select the sensor.
• Figure 3: Closed-loop workflow for the proposed ICLDC. The system integrates an edge server-based LLM module with a safety verifier and an attack detection module. The process begins with the UAV collecting sensory data (e.g., queue length, channel state, battery level) and feedback. This data structures a task description for the LLM, which proposes an optimized schedule via ICL. The safety checker validates this schedule against predefined rules, overriding it if necessary. Concurrently, the attack detection module estimates jailbreaking probability via prompt perplexity analysis, triggering a random safe action upon suspicion. Executed schedules and performance metrics are logged into a feedback loop, enabling continuous LLM adaptation to minimize packet loss.
• Figure 4: Performance analysis of ICLDC in different scenarios with 10 ground sensors concerning the number of Packet Loss (a) The network cost at each time step of ICLDC and baselines: DQN and maximum channel gain, (b) The network cost at each time step of ICLDC with different LLMs (c) Performance with changing the number of ground sensors. (d) Normal and under attack operation of ICLDC, where the jailbreaking attack happens along with MQNS.
• Figure 5: Measured LLM inference latency heatmap superimposed on UAV circular flight path