Predictive maintenance is crucial for reducing operational costs and preventing unexpected failures in industrial systems. Many degradation processes exhibit distinct multi-stage characteristics, where the transition from a normal stage to a defective stage significantly accelerates degradation evolution. Timely detection of such transition is therefore essential for effective maintenance. Consequently, this paper proposes a predictive maintenance framework for multi-stage degradation systems under online defect alarm. Specifically, 1) an online defect alarm algorithm based on cumulative sum (CUSUM) is developed to identify the onset of the defective stage; 2) an empirical Bayes-based parameter updating method is used to estimate the degradation rate in the defective stage; and 3) a predictive maintenance strategy is designed to trigger preventive replacement when the expected remaining useful life falls below a threshold. The proposed framework supports timely and cost-effective maintenance by integrating real-time health assessment with predictive decision-making. The prediction interval and preventive maintenance threshold are jointly optimized to minimize the expected cost rate. Numerical experiments on train bearings demonstrate the effectiveness and superiority of the proposed strategy.

The promise of human-AI teaming lies in humans and AI working together to achieve performance levels neither could accomplish alone. Effective communication between AI and humans is crucial for teamwork, enabling users to efficiently benefit from AI assistance. This paper investigates how AI communication impacts human-AI team performance. We examine AI explanations that convey an awareness of its strengths and limitations. To achieve this, we train a decision tree on the model's mistakes, allowing it to recognize and explain where and why it might err. Through a user study on an income prediction task, we assess the impact of varying levels of information and explanations about AI predictions. Our results show that AI performance insights enhance task performance, and conveying AI awareness of its strengths and weaknesses improves trust calibration. These findings highlight the importance of considering how information delivery influences user trust and reliance in AI-assisted decision-making.

Machine unlearning is widely used to remove hazardous knowledge from large language models. Modern Vision-Language Models (VLMs), however, process both text and visual inputs, raising a fundamental security question: does unlearning in one modality transfer to the other? We present the first systematic, bidirectional study of cross-modal unlearning transfer across three VLM architectures: LLaVA-1.5 (MLP projection), InstructBLIP (Q-Former), and IDEFICS (gated cross-attention). We find that unlearning transfers across modalities, but the transfer is asymmetric and incomplete. In some cases, text unlearning strongly transfers to vision. However, this robustness is not preserved under typographic attacks that manipulate the visual presentation of text. Under such attacks, previously unlearned knowledge can be readily recovered, indicating shallow unlearning. To address the transfer gap and shallow robustness, we propose CrossInf, an influence-guided mitigation strategy. Motivated by the observation that different model components contribute unequally to cross-modal transfer, CrossInf focuses unlearning on transformer blocks that most influence cross-modal generalization. It reduces the transfer gap by more than half in architectures with strong fusion, while preserving model utility. It also improves robustness under typographic attacks, reducing the attack success rate to near zero. We further conduct human evaluation with three annotators (κ=0.77) to validate our findings. Finally, we analyze shallow unlearning using Centered Kernel Alignment (CKA), providing insights into the observed transfer behavior and robustness limitations.

Probability is one of the most useful math skills in daily life. Unfortunately, many students’ understanding of probability is based on preconceptions that are not aligned with established disciplinary meanings. This leads to problems in decision making later in life, where almost everyone, even professional statisticians, suffers from systematic biases in judgments of probability while maintaining strong misconceptions. Artificial Intelligence (AI) offers a setting for probabilistic problem solving that is relevant and meaningful to students, as probability is one of the mathematical concepts that are foundational to AI. Solving problems with AI often involves reasoning under uncertainty, where probability-based concepts can provide tools to explore and reach optimal decisions. For students, AI provides a modern context for connecting probability concepts to real-life situations and provides unique opportunities for transdisciplinary learning that can advance student understanding of both AI systems and probabilistic reasoning. One approach to bring AI to the K-12 classroom that has shown promise in other STEM disciplines is digital game-based learning. Designing game-based environments with AI problem-solving offers a great opportunity to both build on and contribute to the existing knowledge of how to integrate math and AI education in K-12 classrooms through technological innovations. In prior work, we described the design and beta implementation of our educational game, “The 7th Patient”, aimed at teaching probability concepts within an environment that supports the reciprocal development of math and AI skills. This paper presents a more complete implementation of the game, including the full progression through our initial learning objectives, opportunities for transfer learning, and a more interactive environment. We present the games’ mathematical learning objectives, particularly independent and conditional probability, and how we ground those objectives within the application of Bayesian networks: identifying variables, constructing the network with probability tables, performing inference, and decision-theoretic reasoning. We conducted a study where more than 1200 high-school students played the game, and we present our findings on the games’ impact on their knowledge of (and interest in) probability and AI.

Federated learning applied to IoMT can effectively solve the problem of data silos in healthcare, improving healthcare quality and efficiency while ensuring the privacy protection of healthcare data. Since adversaries can track and derive clients' privacy from the shared gradients, federated learning is still exposed to various security and privacy threats. In this paper, we consider two significant concerns regarding the federated learning poisoning attacks in IoMT: (1) how the medical cloud can verify the integrity of the local gradient and (2) how the medical client can verify the correctness of the aggregated results returned. To solve the above problems, we propose two federated learning data integrity verification schemes S-LMI and C-GMI. In the scheme, masking protocol is used to protect the privacy of gradient in federated learning. The homomorphic hash function is explored for both local gradient batch verification and global gradient aggregation verification, ensuring the gradient's integrity. There is no need for data transmission between clients, which can not only protect client privacy but also reduce communication overhead. Our schemes are proved to satisfy unforgeability under the computational Diffie-Hellman problem. The theoretical analysis and extensive experimental results demonstrate that the schemes have high efficiency in the verification process compared to the existing schemes.

The two-degree-of-freedom rotary-linear motor (2DoFRLM), as a promising high-integration mechatronics device, can achieve rotary, linear, and helical motions with a single motor, which is suitable for automation applications fields such as robotic arms, drilling machines, and machine tools. Therefore, it has attracted the attention of both industry and academia, resulting in a large number of different topological designs with their own unique features and operating principles. This article comprehensively reviews the over four decades of literature in this field. The structures of common 2DoFRLMs are introduced, and the motors are classified into three categories based on their winding structure and electromagnetic relationships. Then, the operating mechanisms of different types of motors are systematically analyzed, and their characteristics, applications, and performance are compared. Finally, the recommendations are offered for possible future developments.

RGB photo enhancement is a common computer vision problem with numerous practical application to mobile cameras and on-device photo processing. This creates a need for solutions that are not only performant but are additionally compatible with real mobile AI hardware such as GPUs or NPUs. In this Mobile AI challenge, we address this problem and propose the participants to design efficient RGB image enhancement models that can demonstrate fast inference times on mobile GPUs. For this, the participants were provided with a large-scale DPED dataset consisting of RGB image pairs captured with an old iPhone 3GS smartphone and a professional Canon 70D DSLR camera. The runtime of all models was evaluated on the latest Adreno and Mali GPUs used in Qualcomm and MediaTek chipsets. The proposed solutions are compatible with all recent mobile GPUs, being able to process HD resolution photos under 65 ms in the majority of cases while delivering high-fidelity results. A comprehensive description of the models developed in the challenge is provided in this paper.

Bogie fault diagnosis for bogie is crucial to the safety of rail systems. However, since bogies work under normal states most of the time, the sporadic faulty samples are often submerged in massive normal samples, which are difficult to be distinguished and labeled. Therefore, the labeled training data are always insufficient or even lack of some certain fault states (novel faults), which brings great challenges to fault diagnosis, especially under variable working conditions. Therefore, this paper proposes a new framework named dual-stage manifold preserving mixed supervised learning (d-MMSL) to simultaneously absorb from labeled and unlabeled data effectively. Firstly, manifold similarity (MSLP) is presented to cluster unlabeled samples according to one-off calculation of the manifold similarity. In MSLP, the Best-versus-Second-Best differences and uncertain values are utilized to measure manifold distance and maintain the inherent structure of data. Secondly, Local manifold regularization - broad learning system (LMR-BLS) is presented to o deal with the problem of linear and nonlinear function transformation using simple incremental structure, which could further separate fuzzy sets from MSLP and distinguish the operation conditions of known states accurately. The proposed framework has been verified by a classical dataset and actual vibration data collected from bogies, which achieves a F1-score of 0.99. It is proven that this framework outperforms traditional methods in accuracy and efficiency.

Federated learning (FL) has been shown vulnerable to a new class of adversarial attacks, known as model poisoning attacks (MPA), where one or more malicious clients try to poison the global model by sending carefully crafted local model updates to the central parameter server. Existing defenses that have been fixated on analyzing model parameters show limited effectiveness in detecting such carefully crafted poisonous models. In this work, we propose FLARE, a robust model aggregation mechanism for FL, which is resilient against state-of-the-art MPAs. Instead of solely depending on model parameters, FLARE leverages the penultimate layer representations (PLRs) of the model for characterizing the adversarial influence on each local model update. PLRs demonstrate a better capability to differentiate malicious models from benign ones than model parameter-based solutions. We further propose a trust evaluation method that estimates a trust score for each model update based on pairwise PLR discrepancies among all model updates. Under the assumption that honest clients make up the majority, FLARE assigns a trust score to each model update in a way that those far from the benign cluster are assigned low scores. FLARE then aggregates the model updates weighted by their trust scores and finally updates the global model. Extensive experimental results demonstrate the effectiveness of FLARE in defending FL against various MPAs, including semantic backdoor attacks, trojan backdoor attacks, and untargeted attacks, and safeguarding the accuracy of FL.

Artificial Intelligence (AI) is increasingly integrated into our daily lives. It is critical to prepare future generations with basic knowledge of what AI is, what AI is capable of, and what impact it will have on their lives and career. Informal settings, such as museums, offer unique opportunities to reach out to young learners and importantly, their parents, as they are often the decision-makers for their children's use of AI. In this paper, we discuss the design of an AI-driven exhibit, Virtually Human, that aims to communicate to the public about the capabilities and impact of AI through AI technologies used in virtual humans. The exhibit includes a suite of interactive digital activities and unplugged activities designed to facilitate the AI learning through digital activities. The exhibit aims to engage visitors from 5 to 12 years old, accompanied by their caregivers. The exhibit is open at a local science museum from winter 2023 to summer 2025.

Image super-resolution is a classic computer vision problem with numerous practical applications on mobile and IoT devices. This creates a need for solutions that are not only performant but are additionally compatible with real mobile AI hardware such as neural processing units (NPUs). In this Mobile AI challenge, we address this problem and propose the participants to design efficient quantized deep learning super-resolution models that can demonstrate a real-time performance on mobile NPUs. For this, the participants were provided with the DIV2K dataset and trained quantized models to do an efficient $3 X$ image upscaling. The runtime of all models was evaluated on the Google Tensor NPU present in all recent Google Pixel smartphones. The proposed solutions are fully compatible with all major mobile AI accelerators and are capable of reconstructing Full HD images under 50 ms, delivering highfidelity results. A comprehensive description of the models developed in the challenge is provided in this paper.

As the Metaverse evolves with developments in AI, semantic communication, edge computing, and blockchain, it encounters challenges in adapting to dynamic environments and meeting rising communication and computation needs. In this article, we propose a semantic-aware UAV-based architecture tailored to the dynamic Metaverse that leverages UAV swarms consisting of a collection UAV and edge UAV servers. By mapping different semantic features of Metaverse environments, such as amount of tasks, arrival rate, throughput, and latency requirements, we jointly optimize the mobility, task allocation, and resource allocation in a dynamic Metaverse system. First, a particle swarm optimization-based collection-edge mobility algorithm (PSO-CEMA) is designed to optimize the mobility of UAV servers. Second, to facilitate timely and stable task allocation with reduced complexity, we propose a dual-queue system and a Lyapunov drift function-based dynamic programming task allocation algorithm (LDF-DPTAA). Then, we adopt lifelong learning and design a collection-edge joint training and processing algorithm (LL-CJTPA) to optimize the dynamic allocation of computational resources in the swarm. Finally, we integrate our algorithms into a PSO-LDF-LL algorithm to serve the dynamic Metaverse system. Simulation results show that our approach effectively optimizes UAV servers' positions and task allocation, significantly reduces the training time when facing new tasks, and enhances the stability and efficiency of the network in dynamic settings while reducing congestion.

Backdoor attacks have been extensively explored in recent years which attack deep neural networks (DNNs) by poisoning their training set and causing targeted mis-classification. Research on such attacks is critical for today's widespread applications based on DNNs due to their low-cost and high efficacy. While many backdoor attacks have been proposed, they usually rely on using a static and fixed trigger for attacks, which not only lacks adaptability but also renders them easier to detect. To address such a limitation, we introduce OpenTrigger in this paper, a novel backdoor attack framework employing dynamic triggers for enhancing attack flexibility and robustness. Unlike traditional approaches that rely on a single fixed trigger, our proposed attack learns a generalized consistent feature across a built trigger pool, hence enabling even the use of unseen triggers during testing that differ from those used during training. Extensive experiments across multiple datasets and model architecture confirm the high effectiveness and robustness of OpenTrigger against state-of-the-art and even adaptive backdoor defenses, establishing it as a versatile and practical backdoor attack strategy.

Abstract. Satellite remote sensing provides a unique tool for monitoring dust weather in East Asia in real time and accurately. However, it is still challenging whether these data can effectively and accurately capture the dynamic process of dust weather. Meanwhile, capability and performances of different satellite remote sensing products are not clear in monitoring dust weather. In response to the current problems, this study used PM10 concentration data from environmental monitoring stations to evaluate the continuity, accuracy and stability of five kinds of satellite remote sensing aerosol products (FY4A dust score products (DST) and infrared difference dust index (IDDI), MODIS Aerosol Optical Depth (AOD), Sentinel-5P absorbing aerosol index (AAI) and Himawari-8 AOD) for monitoring and studying dust weather in East Asia. The results showed that the daily spatial distribution of atmospheric dust presented by the five aerosol products had good consistency. In particular, the AAI product was not only better than other aerosol products in depicting the continuity of the spatial distribution of atmospheric dust, but also made up for the inability of other products in obtaining dust information under the clouds. The ground station PM10 data verification showed that the atmospheric dust POCD of MODIS AOD, Himawari-8 AOD, Sentinel-5P AAI, FY4A IDDI and DST products during the entire dust weather process were 91 %, 35.5 %, 24.4 %, 15.8 % and 14.6 respectively. Under the same observation time and space conditions, the atmospheric dust POCD of MODIS AOD, Himawari-8 AOD, FY4A IDDI and DST, and Sentinel-5P AAI products were 85.7 %, 43.8 %, 37.3 %, 37.3 % and 5.6 %, respectively. Overall, the MODIS AOD product performed best in monitoring dust weather in East Asia with high accuracy, and then the Himawari-8 AOD product.

Federated learning is known for its capability to safeguard participants' data privacy. However, recently emerged model inversion attacks (MIAs) have shown that a malicious parameter server can reconstruct individual users' local data samples through model updates. The state-of-the-art attacks either rely on computation-intensive search-based optimization processes to recover each input batch, making scaling difficult, or they involve the malicious parameter server adding extra modules before the global model architecture, rendering the attacks too conspicuous and easily detectable. To overcome these limitations, we propose Scale-MIA, a novel MIA capable of efficiently and accurately recovering training samples of clients from the aggregated updates, even when the system is under the protection of a robust secure aggregation protocol. Unlike existing approaches treating models as black boxes, Scale-MIA recognizes the importance of the intricate architecture and inner workings of machine learning models. It identifies the latent space as the critical layer for breaching privacy and decomposes the complex recovery task into an innovative two-step process to reduce computation complexity. The first step involves reconstructing the latent space representations (LSRs) from the aggregated model updates using a closed-form inversion mechanism, leveraging specially crafted adversarial linear layers. In the second step, the whole input batches are recovered from the LSRs by feeding them into a fine-tuned generative decoder. We implemented Scale-MIA on multiple commonly used machine learning models and conducted comprehensive experiments across various settings. The results demonstrate that Scale-MIA achieves excellent recovery performance on different datasets, exhibiting high reconstruction rates, accuracy, and attack efficiency on a larger scale compared to state-of-the-art MIAs.

The surface-inset permanent magnet synchronous motor (SI-PMSM) with an asymmetric rotor achieves a higher torque with the same amount of permanent magnets when compared to the conventional symmetric SI-PMSM through the strategic offset of magnets on the circumference of the rotor surface. This specific design enables both the reluctance and magnetic torques to achieve maximum values at the same current phase angle simultaneously. However, the asymmetric rotor structure results in an offset between the magnetic d-axis and reluctance d-axis, leading to the inapplicability of the conventional flux weakening (FW) control. In this article, an improved leading-angle FW (ILA-FW) control strategy is proposed to operate the investigated SI-PMSM in the deep FW region and realize wide-speed-range operation. The current operating points of the investigated SI-PMSM under high-speed operation are studied based on the mathematical model where the reference d-axis coincides with the reluctance d-axis. Then, the proposed ILA-FW control strategy is established based on the linearized maximum torque-per-ampere (MTPA) and linearized maximum torque-per-voltage control, which can be divided into MTPA control, partial FW control, and deep FW control. The effectiveness of the proposed control strategy for the investigated SI-PMSM is verified by the prototype experiment and compared with the conventional LA-FW control.