6G networks can support global, ubiquitous and seamless connectivity through the convergence of terrestrial and non-terrestrial networks (NTNs). Unlike terrestrial scenarios, NTNs pose unique challenges including propagation characteristics, latency and mobility, owing to the operations in spaceborne and airborne platforms. To overcome all these technical hurdles, this survey paper presents the use of artificial intelligence (AI) techniques in learning and adapting to the complex NTN environments. We begin by providing an overview of NTNs in the context of 6G, highlighting the potential security and privacy issues. Next, we review the existing AI methods adopted for 6G NTN optimization, starting from machine learning (ML), through deep learning (DL) to deep reinforcement learning (DRL). All these AI techniques have paved the way towards more intelligent network planning, resource allocation (RA), and interference management. Furthermore, we discuss the challenges and opportunities in AI-powered NTN for 6G networks. Finally, we conclude by providing insights and recommendations on the key enabling technologies for future AI-powered 6G NTNs.

Disaster classification and victim detection are two important tasks in enabling efficient rescue operations. In this paper, we propose a multi-task learning (MTL) model which accomplishes these two tasks simultaneously. The idea is to attach one pruned head model to another backbone network. We mathematically pinpoint the optimal branching location and the depth of the pruned head model. Apart from the decoupled task training capability, the MTL model offers lesser memory requirements (12.8 MB saving) and better disaster classification accuracy (1-2% gain), while preserving the same detection performance (0.694 of average precision (AP)), as compared to the traditional method. Such advantages of flexibility, speed and accuracy facilitate the large-scale deployment of Internet of Things (IoT) applications, where we explore the potential of federated learning (FL) and active learning (AL). Given the high ambiguity within disaster images, a modified version of AL-based technique is introduced. For realistic implementation, production-ready OpenFL and OpenVINO tools are adopted to update the global FL model and to optimize the trained model, respectively. Experiment results are promising: the FL-based techniques are comparable to or better than their centralized learning (CL) counterparts. Also, our application portability is demonstrated via different hardware such as CPU and Raspberry Pi.

Federated learning (FL) is a technique that allows multiple clients to collaboratively train a global model without sharing their sensitive and bandwidth-hungry data. This paper presents a joint early client termination and local epoch adjustment for FL. We consider the challenges of heterogeneous Internet of Things (IoT) environments including non-independent and identically distributed (non-IID) data as well as diverse computing and communication capabilities. The goal is to strike the best tradeoff among three conflicting objectives, namely global model accuracy, training latency and communication cost. We first leverage the balanced-MixUp technique to mitigate the influence of non-IID data on the FL convergence rate. A weighted sum optimization problem is then formulated and solved via our proposed FL double deep reinforcement learning (FedDdrl) framework, which outputs a dual action. The former indicates whether a participating FL client is dropped, whereas the latter specifies how long each remaining client needs to complete its local training task. Simulation results show that FedDdrl outperforms the existing FL scheme in terms of overall tradeoff. Specifically, FedDdrl achieves higher model accuracy by about 4% while incurring 30% less latency and communication costs.

The inherent characteristics of Internet of things (IoT) such as low computation power of IoT nodes and transmission reliability of IoT links demand a new paradigm for efficient data processing and dissemination. This is especially true for disaster situations with high possibility of communication breakdowns. On one hand, the concept of artificial intelligence of things (AIoT) has been introduced as a technology to push data storage and computing closer to the network edge. On the other hand, wireless mesh network offers a strong self-healing capability and network robustness against disaster damages. To enable smart disaster monitoring applications, we first implement a lightweight multi-task model that performs joint disaster classification and victim detection. These AI outputs are then wirelessly synchronized via a mesh network solution called NerveNet. All the experiments are conducted in a real urban environment, including static and mobile nodes. Experimental results validate the effectiveness of the proposed solution, where text and images can be synchronized within two minutes across a multi-hop Wi-Fi network. Furthermore, the optimized AI model has ultra-low power consumption around 1.23 W with frames per second (FPS) of 2.01.

The widespread adoption of Unmanned Aerial Vehicles (UAVs) or drones has introduced significant security and privacy challenges, particularly concerning unauthorized drone activities near sensitive areas. To address these concerns, we propose a novel drone detection model, WRN-YOLO, which integrates the Wide Residual Network (WRN) architecture with the You Only Look Once (YOLO) object detection framework. This integration enhances feature extraction capabilities, leading to improved detection accuracy. Through comprehensive ablation studies, we have identified the optimal YOLO variant that synergizes with our backbone modifications, ensuring superior performance in diverse scenarios. Recognizing the complexities of real-world environments, we have also developed a synthetic dataset designed to train our WRN-YOLO. This dataset encompasses a variety of challenging conditions, including intricate backgrounds and the presence of confounding elements, to robustly assess the model's efficacy. Experimental results demonstrate that our method significantly outperforms existing models in accurately detecting drones amidst complex scenes, offering a promising solution for realtime UAV threat mitigation. The proposed approach ranked Top 3 in the 8th WOSDETC Drone-vs-Bird Detection Challenge. Our source code and synthetic dataset are publicly available at https://github.com/yjwong1999/IJCNN2025-DvB.