Open Access

ARTICLE

RSTD-KD: A Task-Cost-Aware Approach for UAV Communication Risk Warning via Knowledge Distillation

Caiyun Li^1,#, Xiaowen Liu^1,#, Binhui Tang^1,*, Tingting Lu², Daibo Xiao³, Li Chen³

1 School of Artificial Intelligence, Hainan Normal University, Haikou, China
2 College of Computer Science, Sichuan University, Chengdu, China
3 Faculty of Humanities and Social Sciences, City University of Macau, Macau, China

* Corresponding Author: Binhui Tang. Email:
# These authors contributed equally to this work

Computers, Materials & Continua 2026, 89(1), 46 https://doi.org/10.32604/cmc.2026.086522

Received 01 June 2026; Accepted 01 July 2026; Issue published 13 August 2026

Abstract

Unmanned aerial vehicle (UAV) communication links in low-altitude operations support mission control, status feedback, and data transmission. Abnormal communication states may affect mission continuity and risk response. However, conventional anomaly detection usually determines only whether a communication anomaly exists, making it difficult to support risk-state assessment and alert-threshold decision-making. To address this problem, this paper proposes Risk Stratification and Task-cost-aware Decision via Knowledge Distillation (RSTD-KD), a task-cost-aware UAV communication risk warning approach that integrates risk stratification, probability calibration, and threshold decision-making. RSTD-KD aggregates fine-grained communication records into window-level behavior samples, constructs low-, medium-, and high-risk reference states, and introduces distillation-guided risk stratification to transfer binary anomaly-boundary information and soft attack-probability supervision into three-level risk-state learning. The estimated attack probabilities are then calibrated for threshold search and cost calculation. Cross-sea logistics and general aviation inspection are used as representative task-cost scenarios to examine the effect of false-positive and false-negative costs on alert-threshold selection. Experiments on the public ECU-IoFT dataset show that RSTD-KD achieves 93.75% Accuracy, 96.44% Balanced Accuracy, and 94.65% Macro-F1. Platt Scaling reduces the Brier Score from 0.0762 to 0.0113. In the additional external-dataset portability evaluation on UAVIDS-2025, RSTD-KD achieves 94.92% Accuracy and 94.38% Macro-F1 across five random seeds, providing further evidence of the portability of RSTD-KD under different UAV communication data granularity and attack taxonomy. These results show that RSTD-KD provides an effective path for converting UAV communication detection outputs into risk states, calibrated attack probabilities, and task-oriented alert decisions.

---

Keywords

---