ACM TIOT launched a special issue on the theme of LLM Empowered Internet of Things, exploring the intersection of Large Language Models (LLMs) and the Internet of Things (IoT). As IoT continues to expand, advanced computational models are increasingly essential for processing and analyzing the massive data generated by interconnected devices. This special issue focuses on how LLMs can enhance IoT systems in several key areas. The second part of this special issue introduces the remaining six accepted papers that spans a board range of IoT scenarios from embedded and cyber-physical systems, human-centered applications, to IoT security.

The Internet of Things (IoT) is rapidly expanding in agriculture, healthcare, and industry, where secure real-time communication is essential. IoT devices often use lightweight protocols such as MQTT due to their limited resources. However, weak security exposes them to cyberattacks, increasing the need for effective Intrusion Detection Systems (IDS). Traditional heuristic methods, including signature-based IDSs, face challenges in MQTT environments due to limited protocol awareness, encrypted payloads, and elusive traffic patterns, leading to high false-positive rates. However, AI-based IDSs improve detection but often suffer from high computational costs and adapt poorly to zero-day attacks. To address these challenges, we propose MGIDS , a hybrid IDS for MQTT IoT networks that handles both encrypted and unencrypted traffic. MGIDS uses unified covariance-distance encoding ( CovDist ) maps flows into quantised bins, which are processed using a dynamic grid-based graph structure ( DYNGrid ). While MGIDS defines the overall framework, DYNGrid serves as the core learning and inference engine. The system adopts a two-stage hierarchical classification strategy, where Stage-1 performs coarse-grained separation, and Stage-2 refines decisions into specific attack types. The proposed method achieves 97.82%, 93.08%, and 97.26% accuracy for known attacks, and 92.80%, 84.86%, and 94.80% for unknown attacks, outperforming existing approaches while demonstrating scalability and robustness.

Detecting earthquakes in seismological time series is a core task in observational seismology, supporting a range of applications from early warning systems to tectonic research. Typically, seismic sensors passively record data and send it to the cloud or edge for integration, storage, and analysis. While this cloud-based approach is effective in urban or well-connected areas, it is impractical in remote, underwater, or underground environments where network infrastructure is unreliable. In such settings, the sensors must operate independently for extended periods while coping with strict constraints on power, memory, and connectivity. To address these challenges, we present LightEQ, a system that combines an efficient data processing pipeline and a lightweight deep-learning model specifically designed for seismic event detection in such environments. LightEQ runs on ultra-low-power microcontrollers with just 100 kB of RAM, enabling real-time, on-device earthquake detection without the need for continuous streaming of raw data to a central location. We evaluate LightEQ against a traditional STA/LTA approach and state-of-the-art (SOTA) machine learning models, using the Stanford Earthquake Dataset. Unlike existing neural network (NN) models, which are too large for microcontrollers, LightEQ is over ten times smaller than most of the SOTA models. Our results demonstrate that communication is the most energy-intensive task in this setting, and that traditional model-driven filters like STA/LTA are inefficient due to their high false positive rate. In contrast, LightEQ improves detection accuracy with NN, providing a more energy-efficient solution by reducing the number of false positives before transmission. Compared to the STA/LTA method alone, LightEQ extends battery life by at least 3-fold by minimizing energy consumption associated with transmitting false positives to the cloud.

Wearable devices are becoming more prevalent in people’s daily lives, particularly in applications such as activity recognition, health monitoring and fitness management. However, the majority of existing wearable devices remain heavily dependent on battery-based power sources, which introduces several practical and sustainability challenges. Frequent battery replacement or recharging imposes inconvenience on users, increases long-term operational costs, and contributes to environmental concerns associated with battery disposal and resource consumption. To address these issues, we present KineticWear, the first battery-free wearable system that utilises kinetic energy harvested from human activities both as the sole energy source and as a sensing signal for on-device human activity recognition (HAR). Based on a careful end-to-end design of all hardware and software components, KineticWear achieves real-time HAR on an ultra low-power microcontroller unit (MCU) including on-board classification and transmission of the inferred activity over a wireless link. Using empirical data, we find that decision tree (DT) and convolutional neural network (CNN) models offer activity recognition accuracies of 87 % and 99.5 % respectively. Systematic real-world experiments demonstrate that KineticWear harvests sufficient energy to operate the wearable device up to 95.2 % of the time, and that the device can infer and report an ongoing activity within 8 seconds using DT classification algorithm, taking three orders of magnitude shorter classification time than CNN. Thus, KineticWear offers significantly enhanced performance compared to state-of-the-art off-device activity recognition systems powered by kinetic energy harvesting.

Autonomous Driving (AD) has garnered significant attention in recent years across multiple domains. Despite notable advancements in algorithms and hardware, large-scale deployment of autonomous vehicles remains constrained by the lack of adequate real-time guarantees. This article comprehensively investigates real-time assurance challenges in AD systems from theoretical and practical perspectives. First, we introduce foundational concepts of real-time systems and analyze common modeling approaches, including the multi-rate DAG and processing chain DAG models. We then delve into the task scheduling and communication mechanisms of three representative middleware systems–ROS2, Cyber, and ERDOS–and the seL4 operating system. Our analysis reveals their underlying design philosophies and optimization strategies for real-time performance. The findings highlight the critical difficulty of meeting stringent real-time requirements in autonomous systems. By offering insights into current limitations and opportunities for improvement, this article aims to establish a deeper understanding of real-time assurance issues and encourage greater focus on system-level guarantees within the AD community.

In the Internet of Vehicles (IoV) systems, recognizing driver emotions is crucial to alleviate dangerous driving behaviors caused by emotional instability. Current research predominantly utilizes multimodal data generated by various types of sensors in IoV systems as input to analyze driver emotion changes using multimodal models. However, existing methods are not enough to fully exploit the advantages of large language models (LLM) in information extraction and multimodal feature fusion, which limits the inference capability of emotion recognition models. Therefore, this article proposes an LLM-auxiliary supervision module, which assists in the training phase through LLM to enhance the performance of multimodal emotion recognition models. Specifically, we designed a label text feature extraction (LTFE) module that employs LLM for text data augmentation and extraction, converting label text into semantically informative feature representations. Additionally, we proposed the label-auxiliary supervision (LAS) strategy, which effectively integrates the LLM label text features learned from the LTFE module with the multimodal emotion recognition model during the training phase to enhance the model’s inference ability. Notably, the LTFE and LAS modules are used only during the training phase, ensuring that the backbone model requires minimal computational resources during inference, making it compatible with the computational constraints of intelligent vehicular devices. Extensive experiments conducted on the PPB-Emo, RAVDESS, and IEMOCAP datasets demonstrate that the proposed method outperforms existing approaches in driver emotion recognition tasks.

Fully decentralized model training for on-road vehicles enables leveraging crowdsourced data without relying on central servers, infrastructure, or persistent Internet connectivity. However, real-world vehicular scenarios pose fundamental challenges to decentralized learning, including highly dynamic network topology, unreliable wireless links, and heterogeneous radio capability constraints. To address these challenges, we propose RoADTrain , a route-assisted decentralized peer model training framework with formal convergence guarantees. In RoADTrain, vehicles share short-horizon route information to estimate inter-vehicle contact duration and link reliability, which are embedded into a base graph. Vehicles then select communication partners by maximizing the algebraic connectivity of the resulting communication graph, thereby accelerating information mixing and model convergence. We further investigate subcarrier reuse and extend the algorithm to three representative radio capability cases. Extensive evaluations show that RoADTrain achieves convergence comparable to communication-unconstrained state-of-the-art approaches while reducing communication overhead by up to 4.5 ×, and significantly outperforms communication-constrained decentralized baselines by up to 29 \(\% \) higher driving success rates in online evaluations. We further empirically demonstrate the robustness of RoADTrain under a range of unexpected effects encountered in complex real-world applications, as well as in a large-scale traffic scenario.

Losing track of reading progress when switching lines can be frustrating. Eye gaze tracking technology offers a potential solution by highlighting read paragraphs, aiding users in avoiding wrong line switches. However, the gap between gaze tracking accuracy (2-3 cm) and text line spacing (3-5 mm) makes direct application impractical. Existing methods leverage the linear reading pattern but fail during jump reading. This paper presents a reading tracking and highlighting system that supports both linear and jump reading. The system leverages the large language model’s contextual perception capability in aiding reading tracking. A reading tracking domain-specific line-gaze alignment opportunity is also exploited to enable dynamic and frequent calibration of the gaze results. Controlled experiments demonstrate reliable linear reading tracking with performance comparable to that of the state-of-the-art linear reading tracking solution, while supporting jump reading tracking with 84% accuracy. Furthermore, real-world field tests with 24 volunteers demonstrated the system’s effectiveness in tracking and highlighting read paragraphs, improving reading efficiency, and enhancing user experience.

The use of low-cost sensors (LCS) in Internet of Things (IoT) networks offers a promising way to improve air quality monitoring. However, there is a major concern regarding their long-term accuracy due to continuous data drift, which requires frequent data recalibration. To address this, we present window-based uncertainty drift detection and recalibration (W-UDDR), a unified system that automates the entire process. Our system uses a Bayesian approach with Gaussian processes (GP) to automatically and accurately detect when sensor output needs to be corrected. To achieve this, the uncertainty of the estimates is quantified using predictive confidence intervals alongside a window system that detects the need for recalibration in real time. We validate our approach using an air quality real-world sensor deployment, systematically assessing key performance metrics such as the frequency of recalibration and the required sample size. Our results show that W-UDDR successfully triggers automatic events when it detects drifts, achieving significant long-term accuracy improvements ranging from 40% to 95%. Hence, this tool provides an automated real-time mechanism to facilitate long-term sensor deployment maintenance.

The rise of wearables such as fitness trackers and smartwatches has increased the need for strong security to protect personal data. Although two-factor authentication methods improve security, they often require additional user input, making them inconvenient. Recently, hardware flaws in accelerometers and WiFi interfaces have been leveraged to create low-effort two-factor authentication methods. However, these hardware-based device credentials are static, necessitating device replacement if the credentials are compromised. In this study, we introduce an innovative device authentication system that identifies wearables using vibration-based credentials. By utilizing built-in vibration motors and motion sensors (i.e., accelerometers and gyroscopes), our system establishes a unique communication channel to capture the distinct characteristics of each device. Unlike existing methods, our vibration-based credentials are reprogrammable and user-friendly. We develop advanced data processing techniques to minimize the impact of noise, body motion artifacts, and wearing position. We design a lightweight convolutional neural network for feature extraction and device authentication, with a majority vote mechanism to improve identification robustness. Extensive experiments with five different smartwatches demonstrate that our system achieves an average precision of 98% and a recall of 94% under various attacks, demonstrating that including gyroscope data significantly improves performance across different wearing poses and watch orientations.