Predicting spatiotemporal patterns is critical for traffic flow forecasting in Intelligent Transportation Systems (ITS), as accurate predictions can significantly enhance traffic management and decision-making. Recent data-driven approaches, particularly Graph Neural Networks (GNNs) integrated with physics-informed partial differential equations (PDEs), have demonstrated strong capability in modeling complex traffic dynamics. However, these models may still produce overconfident yet incorrect predictions, which can be particularly problematic in safety-critical scenarios. Uncertainty quantification (UQ) enhances the reliability of predictions and the practical application of NNs by estimating prediction confidence, not just accuracy. Despite its importance, existing UQ methods in traffic forecasting are primarily developed for purely data-driven models, and the role of physics-informed modeling in uncertainty estimation remains insufficiently understood. In this paper, we bridge this gap by integrating multiple UQ baseline methods with physics-informed modeling to systematically investigate how physical constraints influence uncertainty estimation in traffic forecasting. We further propose a physics-informed loss function that enhances the model's ability to capture physically consistent dynamics and improves the calibration and reliability of uncertainty estimates. In addition, we evaluate robustness under both noisy and adversarial perturbations, showing that our approach yields more stable predictions and uncertainty estimates under distribution shifts. Extensive experiments on real-world traffic datasets demonstrate that our approach improves both prediction accuracy and uncertainty quality, achieving up to 14.8% improvement in short-term and 8.7% in long-term traffic speed prediction errors, while providing better-calibrated and more robust uncertainty estimates.
The newly proposed multimodal transformer architecture offers a new paradigm for UAV detection and aerial object recognition. It introduces an innovative way of feeding multiple data streams, such as audio, infrared video, RGB video, and radar, into the architecture for processing, using independent modalities. The unique features of each modality are attached and processed together in the architecture, where the features are then exposed to the multimodal transformer for classification. Thus, all complementary information can be pooled within the integration framework to allow the model discrimination of any drone target under outdoor conditions from other aerial objects such as birds, helicopters, and airplanes. These methodologies are expected to outperform traditional single-modality systems by improving detection accuracy through class balancing and addressing modality-specific limitations. The proposed model has been further tested through various experiments to evaluate its robustness under conditions such as missing entries, corrupted data, and synthetic inputs. The results suggest that it has strong potential to serve as a benchmark in UAV detection. Thus, this work takes part of an emerging body of sensor fusion and deep learning-related research, demonstrating the potential of multimodal data in real-world detection problems.
The development of embodied AI (EAI) critically depends on efficient data acquisition, yet faces persistent challenges including high costs, limited training scenarios, and lack of standardized datasets. We present AIRSPEED, an open source data production platform designed to address these bottlenecks through three core innovations. First, AIRSPEED achieves hardware–software decoupling via unified robot and simulation interfaces, enabling seamless integration with diverse data collection devices and simulation platforms. Second, it supports comprehensive data production methods spanning teleoperation and teaching approaches, as well as synthetic data generation through data synthesis and virtual teleoperation. Third, AIRSPEED automates pyramid-structured dataset construction compatible with both HDF5 and LeRobot formats, significantly reducing manual overhead. Experimental validation demonstrates substantial efficiency gains, achieving up to 35.6× acceleration in dataset construction and 6.0× overall speedup compared to manual workflows. With end-to-end latency as low as 3 ms and compression throughput exceeding 296 MB/s, AIRSPEED establishes a scalable foundation for EAI data production. AISPEED is open sourced on this website: URL .
The deep convergence of Information Technology (IT) and Operational Technology (OT) exposes Industrial Internet of Things (IIoT) systems to complex cross-layer attacks. However, traditional defense methods are mostly designed for a single domain and rely on static or rule-based mechanisms, making them ineffective in capturing cross-layer attack evolution and coordinating adaptive responses. Therefore, this article proposes a Situation-Based Hierarchical Multi-Agent Reinforcement Learning (Situ-HMARL) adaptive defense framework, which formulates defense strategies based on real-time global industrial situations and proactively enhances the resilience of IIoT systems. Firstly, a three-stage industrial situational awareness architecture is proposed to continuously fuse heterogeneous data from the IT and OT layers into a structured Global Industrial Situation Vector (GISV), which serves as a unified global observation space for defense decision-making. Secondly, a lightweight Moving Target Defense (MTD) mechanism is designed to adaptively trigger IP-hopping based on the global industrial situation, thereby reducing overhead while preserving system availability. On this basis, a two-level Hierarchical Multi-Agent Reinforcement Learning (HMARL) framework is developed to decouple perception, decision-making, and enforcement. Low-level agents perform real-time local situational perception and execute defense actions, while a high-level agent reasons over the global industrial situation to generate coordinated defense strategies that minimize system losses. Extensive experiments conducted on the Cyber Operations Research Gym (CybORG) platform validate that the proposed framework effectively mitigates cross-layer attacks and significantly improves the availability of IIoT systems.
This paper presents an innovative approach that integrates Embodied Artificial Intelligence (EAI) into Virtual Reality (VR) by developing trustworthy intelligent avatars capable of adapting to users’ physiological states and behavioral nuances, thereby enhancing trustworthiness. We introduce ”SmartSimVR,” a versatile framework that leverages Artificial Intelligence (AI) and sensor fusion to create these intelligent avatars. The framework facilitates real-time learning and adaptation within a closed loop. To validate our framework, we implemented SmartSimVR in a driving simulator, where the car functions as a non-humanoid avatar representing the user. This human-centered intelligence aims to mitigate cybersickness, a key factor in improving user experience. The trustworthy intelligent avatar continuously learns from both objective data—encompassing multimodal physiological and behavioral data—and subjective feedback, adjusting its behavior in real-time if it infers cybersickness symptoms by modifying car acceleration after a dynamic threshold time frame. Our experiment involved 30 subjects participating in two sessions, allowing for effective comparison of results. Analysis conducted in the time-frequency domain, using spectrogram power and wavelet energy metrics, identified crucial objective measures—such as eye velocity, eye angular velocity, and electrodermal activity (EDA)—as sensitive indicators of cybersickness. Notably, we observed significant reductions in these measures among users interacting with the avatar, reinforcing its effectiveness in minimizing cybersickness. Additionally, subjective assessments using the Simulator Sickness Questionnaire (SSQ) revealed a substantial 44.3% reduction in symptoms when participants engaged with the avatar. Strong positive correlations between the wavelet energy of these indicators and SSQ scores further support our findings. Together, both subjective and objective results provide compelling evidence that the avatar can significantly enhance user experiences in VR environments.
The rapid expansion of connected devices has made them prime targets for cyberattacks. To address these threats, deep learning-based, data-driven intrusion detection systems (IDS) have emerged as powerful tools for detecting and mitigating such attacks. These IDSs analyze network traffic to identify unusual patterns and anomalies that may indicate potential security breaches. However, prior research has shown that deep learning models are vulnerable to backdoor attacks, where attackers inject triggers into the model to manipulate its behavior and cause misclassifications of network traffic. In this article, we explore the susceptibility of deep learning-based IDS systems to backdoor attacks in the context of network traffic analysis. We introduce PCAP - Backdoor , a novel technique that facilitates backdoor attacks on PCAP datasets. Our experiments on real-world Cyber-Physical Systems (CPS) and Internet of Things (IoT) network traffic datasets demonstrate that attackers can effectively backdoor a model by poisoning as little as 2% or less of the entire training dataset. Moreover, we show that an attacker can introduce a trigger into benign traffic during model training yet cause the backdoored model to misclassify malicious traffic when the trigger is present. Finally, we highlight the difficulty of detecting this trigger-based backdoor, even when using existing backdoor defense techniques.
This is a corrigendum for the article “Quantitative Measurement of Cyber Resilience: Modeling and Experimentation” published in ACM Trans. Cyber-Phys. Syst. 9, 1, Article 1 (January 2025), 25 pages.
Bayesian Optimization (BO) is a widely used technique for optimizing black-box functions, whose mathematical form is unknown and can only be evaluated through costly simulations. In this work, we focus on optimizing functions that are the \(\min/\max\) of several components, where each component is typically a non-linear, non-convex black-box function. Traditional BO approaches may suffer from the masking effect , where components with lower mean values are sampled more, even if they do not contain the global optimum. In our prior work, Minimum Bayesian Optimization ( minBO ) addressed this issue by proposing a sampling approach that uses a surrogate for each component and samples based on predicted improvement across all components. While effective, minBO treats each component independently and does not capture potential dependencies between components. Additionally, estimating surrogates for each component can limit the scalability of the approach. We introduce Conjunctive Bayesian Optimization ( conBO ), a novel approach that overcomes these limitations. We propose a paired sampling algorithm ( conBO-PS ) that considers dependencies between components by analyzing all pairs of functions and estimating the distribution of the minimum of two functions. While conBO-PS accounts for dependencies, it is computationally expensive. To improve efficiency, we introduce conBO large-scale ( conBO-LS ), which adapts conBO-PS by considering a subset of components chosen based on their potential impact, allowing users to control computational effort. We evaluate the performance of these algorithms on non-linear synthetic functions and also compare them to state-of-the-art methods in the context of falsifying conjunctive safety requirements for Cyber-Physical Systems (CPS). A conjunctive safety requirement refers to a set of safety conditions (requirements) tested together, where if at-least one condition is falsified, the entire conjunctive requirement is considered falsified. In fact, in such context the function to be minimized is the minimum of several sub-components. Results show that conBO-PS and conBO-LS outperform existing approaches, offering better solution quality and computational efficiency. In the CPS application, the proposed approaches achieve faster falsification and improved falsification rates across all benchmarks.
Autonomous cyber-physical systems depend on high-resolution sensors that generate multi-gigabyte-per-second data streams for accurate environmental perception and system safety. However, existing publish-subscribe middleware frameworks, such as ROS 2 with DDS, are optimized for small data objects and face significant latency and overhead challenges when handling large sensor data between distributed application nodes. Recognizing the need for more efficient data management, we propose a new companion middleware that prioritizes application-specific data relevance, enabling selective communication of critical information while reducing the burden on communication resources and maintaining interoperability with state-of-the-art publish-subscribe middleware. This companion middleware enables timely and effective sharing of sensor data by focusing on regions of interest relevant to specific tasks, such as traffic light detection in driving scenarios. Experimental evaluations of our open source implementation of this companion middleware on a Linux platform demonstrate that our protocol integrates efficiently with ROS 2, significantly enhancing data management and communication efficiency.
This work addresses congestion attacks on prioritization and preemption signal applications (PPSA), an important variant of cooperative intelligent transport systems (C-ITS). In PPSA, priority vehicles, such as buses and ambulances, can request traffic signal adjustments in their favor to reduce travel time. These systems rely on vehicle-to-everything (V2X) communications, which makes them vulnerable to congestion attacks. When under attack, these systems experience disruptions in information exchange between vehicles and infrastructure, leading to traffic congestion and safety hazards. Nevertheless, existing solutions do not address the attacks at a system level and within the PPSA context. Our proposal presents a software framework designed to enhance both sustainability and safety during congestion attacks. Specifically, it introduces a whitelist-based traffic-filtering mechanism that preserves system sustainability by ensuring only legitimate priority requests are processed. Additionally, a monitoring mechanism evaluates the PPSA service rate at runtime to identify and address potential safety risks arising from insufficient service rates. Moreover, we analyze the limitations of a machine learning-based defense method and highlight risks of potential service unavailability due to false positives. We believe that this research contributes to the development of a secure and safe traffic management system for smart cities.
In modern vehicular systems, substantial computation is required for functions such as high-definition mapping, object detection, and driving assistance. While onboard computing offers low-latency access, it is sometimes insufficient in capacity. Cloud computing provides scalability but suffers from latency issues. Edge computing offers a balance of low-latency, localized service delivery, and enhanced computational capacity, but it demands sophisticated allocation and delivery mechanisms. In this paper, we address the service allocation problem in edge-assisted vehicular networks under realistic assumptions of variable but bounded vehicle velocities. We design efficient computation allocation algorithms and introduce a speed-adjustment suggestion mechanism for vehicles not immediately assigned resources, allowing them to slow down and increase future allocation chances. Moreover, we propose a dedicated scheduling strategy tailored to normal traffic and platoon-based traffic, leveraging their structural properties. To mitigate computational overhead, we develop a novel and scalable approximation technique. Experimental evaluations across various traffic scenarios demonstrate the effectiveness, scalability, and real-time applicability of the proposed methods. Our code is available at https://github.com/AAAA-source/Edge-Assisted-Service-Allocation-for-Connected-Vehicles-with-Variable-Velocities .
Modern publish-subscribe systems are increasingly deployed in applications such as smart grids, industrial IoT, and smart homes, where brokers perform not only message forwarding but also in-broker computation. This shift from receive-forward (RF) to receive-compute-forward (RCF) models introduces new privacy and trust concerns, as centralized brokers gain access to sensitive data. In this article, we present a distributed and privacy-preserving RCF ( \(\textsf{DRCF}\) ) framework that replaces the trust assumption on a single centralized broker with multiple semi-honest brokers. Our system integrates distributed authenticated encryption, threshold signatures, and 2-party secure computation realized using homomorphic encryption and garbled circuits to protect both communication and computation. We formally prove the security of \(\textsf{DRCF}\) in the semi-honest adversarial model. We implement \(\textsf{DRCF}\) on top of the Mosquitto broker and evaluate it on real-world IoT workloads, including anomaly detection, load forecasting, and electricity price prediction. Experimental results demonstrate that \(\textsf{DRCF}\) achieves scalability and efficiency while preserving privacy, confidentiality, and integrity in modern publish-subscribe systems.
Federated Learning (FL), a decentralized Machine Learning (ML) approach, allows Wireless Body Area Network (WBAN) users to collaboratively train models while maintaining the privacy of their health data. With the rise of ML-powered smart healthcare applications and increasing demand for diverse services, the simultaneous training of multiple FL models using data from WBANs is becoming feasible. However, managing multiple FL models with distinct learning objectives presents challenges in designing incentives and selecting suitable users—an area that remains largely unexplored. Thus, this article introduces an auction-based incentive mechanism and WBAN users selection framework to enable the parallel training of multiple FL models, ensuring privacy of data. An optimization problem is formulated to maximize system utility, incorporating a cost model that includes data collection, computation, communication, and privacy. The proposed auction-based algorithm integrates factors such as local model accuracy, user reputation, and data volume to solve this problem efficiently. Simulations and real-world health data analysis demonstrate that this approach improves average utility by 15.9% and 18.08% compared to state-of-the-art methods.
Socially Assistive Robots (SARs) would benefit from exhibiting a higher level of embodied artificial intelligence by being able to engage naturally with people, acquiring an understanding of their operating environments and contextually relevant information. Deep neural networks (DNNs) show promise in enabling SARs to recognize a variety of stimuli. Yet, deploying them in real-world settings presents challenges such as adapting to evolving environments, avoiding the need for specialized datasets, and ensuring privacy by minimizing cloud processing. This paper introduces Human-Interactive Robot Learning with Feedback (HIRL-F), an approach where SARs can learn from human feedback. HIRL-F leverages DNN embeddings as invariant representations from which to confirm and designate percepts. Using memory modules, HIRL-F validates percepts through human input, and establishes associations based on prior experiences, synchronicity, and spatial proximity. The approach is evaluated using the robot T-Top, an interactive tabletop experimental platform designed for onboard DNN processing. The paper presents the implementation along with controlled and open-environment evaluations, to demonstrate the feasibility and potential of HIRL-F for enhancing cognitive capabilities of interactive robots through real-time, in situ learning in open settings.
The increasing integration of smart grids within the Internet of Things (IoT) ecosystem requires the implementation of robust security measures due to vulnerabilities brought about by pervasive connectivity. Previous smart grid security protocols are based on classical cryptographic algorithms, which are susceptible to significant threats from emerging quantum computers. Consequently, the adoption of quantum-resistant solutions is imperative for the long-term security of smart grids. This paper presents a novel protocol for authentication and secret key establishment in smart grids, utilizing post-quantum cryptography (PQC) algorithms. Our proposed protocol employs the FALCON digital signature algorithm for integrity verification and the CRYSTALS-Kyber key encapsulation mechanism (KEM) for secret key establishment, providing robust protection against quantum attacks. Our optimized implementation of the proposed protocol on a graphics processing unit (GPU) targets grid security module (GSM)/gateway-side deployments and demonstrates scalability under large numbers of concurrent authentication requests. Experimental validation demonstrates the proposed protocol's ability to effectively manage numerous concurrent authentication requests, ensuring secure and efficient communication within smart grid networks.
Safe Reinforcement Learning (RL) has been applied to synthesize control policies that maximize task rewards while adhering to safety constraints within simulated secure cyber-physical systems. However, the vulnerability of safe RL to adversarial attacks remains largely unexplored. We argue that understanding the safety vulnerabilities of learned control policies is crucial for ensuring true safety in real-world scenarios. To address this gap, we first formally define the safe RL problem with formal language (signal temporal logic) and demonstrate that even optimal policies are susceptible to observation perturbations. We then introduce novel safety violation attacks that exploit adversarial models trained with reversed safety constraints to induce unsafe behaviors. Lastly, through both theoretical analysis and experimental results, we demonstrate that our approach is more effective at violating safety constraints than existing adversarial RL methods, which primarily focus on reducing task rewards rather than compromising safety.
Roadside cameras effectively enhance the perception capabilities of embodied artificial intelligence systems such as vehicles by compensating for the limitations of vehicle-mounted cameras, which are prone to occlusion and have a limited sensing range, thereby improving the safety of autonomous vehicles. However, existing object detection systems often encounter perception errors when handling comprehensive viewpoint noise in roadside scenes, as well as variations in traffic flow, lighting conditions, and camera poses. This makes it challenging for them to perform robustly in complex road environments. To address these issues, we propose \(\mathrm{R^{2}MOAG}\) , a highly robust monocular 3D object detection method for roadside systems, based on ground perception embedding and heterogeneous visual tokens. The proposed method extracts detailed road information through ground plane equations and utilizes heterogeneous visual tokens to focus on foreground features. By integrating low-dimensional ground information with high-dimensional visual features, the model is provided with clear and rich cues for object detection, significantly enhancing its stability. We conducted extensive experiments on the widely recognized roadside datasets DAIR-V2X-I and Rope3D. The results show that, in terms of overall performance, the proposed model achieved a 4.65% and 4.26% improvement in the \(AP_{3D}|_{R40}\) metric for the vehicle category on these two datasets, respectively. Moreover, the model maintained stable recognition performance across various road scenarios and camera poses, demonstrating exceptional robustness.
The integration of Artificial Intelligence (AI) into software testing has emerged as a promising approach to addressing the limitations of conventional methods, particularly in safety-critical domains such as automotive and aerospace. Traditional software testing is often manual, resource-intensive, and prone to incomplete coverage. In contrast, AI-based techniques have been applied to automate test case generation, optimise testing processes, and improve defect prediction. This study extends these approaches to the context of Embodied AI, where the reliability of systems that perceive and act within complex environments is a primary concern. Autonomous satellites are considered as representative embodied agents, for which the robustness of the perception–action loop is mission-critical. The paper presents two main contributions. First, it reviews the state of the art in the application of AI techniques to software testing. Second, it demonstrates the practical viability of these techniques through a proof of concept (PoC) in the satellite domain. In this PoC, generative AI is used to create synthetic datasets simulating adverse atmospheric conditions, including varying levels of cloud opacity. These datasets are used to evaluate and improve an on-board object detection model. Experimental results indicate that this approach reveals limitations of the baseline model and, following retraining, enhances robustness under challenging conditions while maintaining performance in clear scenarios. The findings demonstrate the potential of AI-based testing methodologies to support the validation and improvement of mission-critical embodied systems.
