The growing ubiquity of cyberphysical systems (CPSs) embedded in the Internet of Things (IoT)–edge–cloud (IEC) continuum is transforming how data-driven applications are deployed and operated. Existing approaches to fault detection, diagnosis, and healing in such CPS deployments predominantly rely on centralized or statically supervised machine learning models. Due to the growing complexity of CPS systems, such as autonomous vehicles and smart cities, which require time-sensitive responses and utilize resource-constrained IoT and edge devices, fault detection presents several formidable research challenges. In this regard, multiagentic artificial intelligence, coupled with lifelong learning, offers a promising foundation. Despite its promise, realizing such decentralized and intelligent fault-management paradigms becomes not just beneficial, but necessary.

Numerous deep learning-based works focusing on 3D semantic segmentation have been proposed and have achieved impressive performance. However, due to the catastrophic forgetting, existing methods will degrade dramatically in a real-world scenario where new 3D semantic categories are arriving continually. Straightforwardly applying typical class-incremental learning methods on 3D data even aggravates forgetting due to the irregular and noisy geometric structure. Aiming to address this realistic challenge, from the perspective of capturing local topological characteristics and mitigating global semantic shift, we propose a unified framework named Local topological Alignment and Global semantic Deconstruction (LAGD) to incrementally learn semantic knowledge of novel 3D categories while maintaining performance on previously learned knowledge. Specifically, we develop a novel Interaction Topological-aware Alignment (ITA) to maintain the learned knowledge efficiently by capturing the local geometric characteristics with interacted adjacent state-specific knowledge. Besides, to mitigate the forgetting caused by the global semantic shift, we deconstruct the logits into positive and negative parts which are distilled separately, achieving an elaborate distillation process in terms of Semantic-knowledge Deconstruction Distillation (SDD). With the cooperation of ITA and SDD, LAGD achieves a sota performance, especially in the long-term incremental learning scenario. Extensive experimental results illustrate the superiority of our proposed LAGD.

Connected and autonomous vehicles (CAVs) are becoming increasingly common, and their popularity is expected to increase further, especially due to the ease and convenience they provide. An important component of CAVs is the Electronic Control Unit or the ECU; these ECUs are connected to each other through the Controller Area Network or CAN bus. The increasing use of CAVs has significantly expanded its threat landscape with several security vulnerabilities that can be severely damaging, especially when the ECUs or CAN are compromised. Existing security mechanisms are largely designed for known vulnerabilities, leaving the potential for unknown attacks high. In order to address this gap in detecting previously unknown attacks, we propose a novel autoencoder and payload fragmentation-based in-vehicle unknown intrusion method. For the precise prediction of attack and sub-attack, the data payload is fragmented into one-byte fragments, and the prediction dependency is analyzed for each fragment. Extensive evaluations demonstrate that our proposed method shows improved performance with an attack classification accuracy of 99.88% and sub-attack accuracy of 99.98%.

This research delves into the application of Federated Learning (FL) models for detecting fraud across different financial bodies. FL facilitates decentralized training of models using local data, ensuring privacy, crucial for handling sensitive financial data. The comparison involves three machine learning models - Artificial Neural Networks (ANN), Random Forest (RF), and Convolutional Neural Networks (CNN) - to assess their efficacy in the FL context. While ANN and CNN demonstrate strong capacity in identifying complex fraud patterns, their communication efficiency and overfitting challenges are significant. In contrast, RF offers more robustness to Non-independent and Identically Distributed (non-IID) data and is less prone to overfitting, though it poses communication overhead issues. This paper also highlights the challenges of FL in fraud detection, including data heterogeneity, communication costs, and security risks. This paper proposed future research directions, emphasizing model personalization, communication optimization, and advanced privacy-preserving techniques. By addressing these challenges, FL can offer scalable, secure solutions for real-time fraud detection, ensuring the protection of sensitive financial data while enhancing detection accuracy across diverse data sources.

We introduce a rehearsal-free federated domain incremental learning framework, RefFiL, based on a global prompt-sharing paradigm to alleviate catastrophic forgetting challenges in federated domain-incremental learning, where unseen domains are continually learned. Typical methods for mitigating forgetting, such as the use of additional datasets and the retention of private data from earlier tasks, are not viable in federated learning (FL) due to devices' limited resources. Our method, RefFiL, addresses this by learning domain-invariant knowledge and incorporating various domain-specific prompts from the domains represented by different FL participants. A key feature of RefFiL is the generation of local fine-grained prompts by our domain adaptive prompt generator, which effectively learns from local domain knowledge while maintaining distinctive boundaries on a global scale. We also introduce a domain-specific prompt contrastive learning loss that differentiates between locally generated prompts and those from other domains, enhancing RefFiL's precision and effectiveness. Compared to existing methods, RefFiL significantly alleviates catastrophic forgetting without requiring extra memory space, making it ideal for privacy-sensitive and resource-constrained devices.

We propose Exemplar-Condensed federated class-incremental learning (ECoral) to distil the training characteristics of real images from streaming data into informative rehearsal exemplars. The proposed method eliminates the limitations of exemplar selection in replay-based approaches for mitigating catastrophic forgetting in federated continual learning (FCL). The limitations particularly related to the heterogeneity of information density of each summarized data. Our approach maintains the consistency of training gradients and the relationship to past tasks for the summarized exemplars to represent the streaming data compared to the original images effectively. Additionally, our approach reduces the information-level heterogeneity of the summarized data by inter-client sharing of the disentanglement generative model. Extensive experiments show that our ECoral outperforms several state-of-the-art methods and can be seamlessly integrated with many existing approaches to enhance performance.

Hierarchical federated learning (HFL) is a promising distributed deep learning model training paradigm, but it has crucial security concerns arising from adversarial attacks. This research investigates and assesses the security of HFL using a novel methodology by focusing on its resilience against inference-time and training-time adversarial attacks. Through a series of extensive experiments across diverse datasets and attack scenarios, we uncover that HFL demonstrates robustness against untargeted training-time attacks due to its hierarchical structure. However, targeted attacks, particularly backdoor attacks, exploit this architecture, especially when malicious clients are positioned in the over-lapping coverage areas of edge servers. Consequently, HFL shows a dual nature in its resilience, showcasing its capability to recover from attacks thanks to its hierarchical aggregation that strengthens its suitability for adversarial training, thereby reinforcing its resistance against inference-time attacks. These insights underscore the necessity for balanced security strategies in HFL systems, leveraging their inherent strengths while effectively mitigating vulnerabilities.

An emerging evacuation path planning technique that uses Low Power Wide Area Networks (LPWAN) to enable real-time danger prediction and user-oriented path planning can ensure the safe and timely navigation of evacuees in complex scenarios such as cruise ships. However, most existing LPWANbased evacuation models assume pedestrians’ walking speed remains constant and ignore crowd congestion in corridors before exits, which is not appropriate for rocking ships. To overcome these issues, this paper proposes a congestion-relived guiding framework with dedicated path planning for emergency evacuation on passenger ships. The basic idea is to averagely minimize the total evacuation time while meeting the deadline for ship capsizing under all circumstances by selecting uncrowded paths for each passenger individually. First, we use probability distributions rather than constant numbers to represent walking time (also called delay) along passageways. A worst-case delay bound with a high level of trustworthiness is also estimated for each passageway under the boundary condition of ship capsizing. Next, we predict the congestion of corridors by modeling the spatiotemporal movement of passengers, and then distribute evacuation loads evenly among corridors to alleviate the congestion. The total expected evacuation time of all corridors is finally minimized based on the delay probability distribution and estimated congestion, and the deadline for ship evacuation under all circumstances is met with the worst-case delay bound. Simulation results show that our approach significantly reduces the total escaping time of crowd evacuation by 45% and 34% while improving the navigation success ratio by more than 20% and 80% compared with the state-of-the-art emergency evacuation systems, namely the look-up table guiding scheme and the group-based guiding evacuation scheme, respectively.

With impressive achievements made, artificial intelligence is on the path forward to artificial general intelligence. Sora, developed by OpenAI, which is capable of minute-level world-simulative abilities can be considered as a milestone on this developmental path. However, despite its notable successes, Sora still encounters various obstacles that need to be resolved. In this survey, we embark from the perspective of disassembling Sora in text-to-video generation, and conducting a comprehensive review of literature, trying to answer the question, \textit{From Sora What We Can See}. Specifically, after basic preliminaries regarding the general algorithms are introduced, the literature is categorized from three mutually perpendicular dimensions: evolutionary generators, excellent pursuit, and realistic panorama. Subsequently, the widely used datasets and metrics are organized in detail. Last but more importantly, we identify several challenges and open problems in this domain and propose potential future directions for research and development.

Class incremental learning (CIL) strives to emulate the human cognitive process of continuously learning and adapting to new tasks while retaining knowledge from past experiences. Despite significant advancements in this field, Transformer-based models have not fully leveraged the potential of attention mechanisms to balance the transferable knowledge between tokens and the associated information. This paper addresses this gap by using a dual variational knowledge attention (DVKA) mechanism within a Transformer-based encoder-decoder framework, tailored for CIL. DVKA mechanism aims to manage the information flow through the attention maps, ensuring a balanced representation of all classes, and mitigating the risk of information dilution as new classes are incrementally introduced. This method, leverage the information bottleneck and mutual information principle, selectively filters less relevant information, directing the model's focus towards the most significant details for each class. The DVKA is designed with two distinct attentions: one focused on the feature level and the other on the token dimension. The feature-focused attention aims to purify the complex nature of various classification tasks, ensuring a comprehensive representation of both old and new tasks. The token-focused attention mechanism highlights specific tokens, facilitating local discrimination among disparate patches and fostering global coordination for a spectrum of task tokens. Our work is a major stride towards improving transformer models for class incremental learning, presenting a theoretical rationale and effective experimental results on three widely-used datasets.

Researchers have recently focused on multimodal emotion recognition, but issues persist in recognizing emotions in multi-party dialogue scenarios. Most studies have only used text and audio modality, ignoring the video modality. To address this, we propose M2ER, a multimodal emotion recognition scheme based on multi-party dialogue scenarios. Addressing the issue of multiple faces appearing in the same frame of the video modality, M2ER introduces a method using multi-face localization for speaker recognition to eliminate the interference of non-speakers. The attention mechanism is used to fuse and classify different modalities. We conducted extensive experiments in unimodal and multimodal fusion using the multi-party dialogue dataset MELD. The results show that M2ER achieves superior emotion recognition in both text and audio modalities compared to the baseline model. The proposed method using speaker recognition in the video modality improves emotion recognition performance by 6.58% compared to the method without speaker recognition. In addition, the multimodal fusion based on the attention mechanism also outperforms the baseline fusion model.

Summary Early warning systems (EWS) for electrical grid infrastructure have played a significant role in the efficient management of electricity supply in natural hazard prone areas. Modern EWS rely on scientific methods to analyze a variety of Earth Observation and ancillary data provided by multiple and heterogeneous data sources for the monitoring of electrical grid infrastructure. Furthermore, through cooperation, EWS for natural hazards contribute to monitoring by reporting hazard events that are associated with a particular electrical grid network. Additionally, sophisticated domain knowledge of natural hazards and electrical grid is also required to enable dynamic and timely decision‐making about the management of electrical grid infrastructure in serious hazards. In this paper, we propose a data integration and analytics system that enables an interaction between natural hazard EWS and electrical grid EWS to contribute to electrical grid network monitoring and support decision‐making for electrical grid infrastructure management. We prototype the system using landslides as an example natural hazard for the grid infrastructure monitoring. Essentially, the system consists of background knowledge about landslides as well as information about data sources to facilitate the process of data integration and analysis. Using the knowledge modeled, the prototype system can report the occurrence of landslides and suggest potential data sources for the electrical grid network monitoring.

A platform for orchestrating networked machine learning (ML) applications over distributed environments is described. ML applications are transformed into automated pipelines that manage the whole application lifecycle and production-grade implementations are automatically constructed. We present AUTOSTEER, a software platform that can deploy ML applications on various hardware resources—interconnected using heterogeneous network resources—across cloud and edge devices. Device placement optimization and model adaptation are used as control actions to support application requirements and maximize the performance of ML model execution over heterogeneous computing resources. The performance of deployed applications is continually monitored at runtime to overcome performance degradation due to incorrect application parameter settings or model decay. Three real-world applications are used to demonstrate how AUTOSTEER can support application deployment and runtime performance guarantees.

Social media has played a significant role in disaster management, as it enables the general public to contribute to the monitoring of disasters by reporting incidents related to disaster events. However, the vast volume and wide variety of generated social media data create an obstacle in disaster management by limiting the availability of actionable information from social media. Several approaches have therefore been proposed in the literature to cope with the challenges of social media data for disaster management. To the best of our knowledge, there is no published literature on social media data management and analysis that identifies the research problems and provides a research taxonomy for the classification of the common research issues. In this paper, we provide a survey of how social media data contribute to disaster management and the methodologies for social media data management and analysis in disaster management. This survey includes the methodologies for social media data classification and event detection as well as spatial and temporal information extraction. Furthermore, a taxonomy of the research dimensions of social media data management and analysis for disaster management is also proposed, which is then applied to a survey of existing literature and to discuss the core advantages and disadvantages of the various methodologies.

We demonstrate AHOM, a system that can Actively Observe Hazards via Monitoring Social Media Streams. AHOM proposes an active way to include the human in the loop of hazard information ac-quisition for social media. Different from state of the art, it supports bi-directional interaction between social media data processing system and social media users, which leads to the establishment of deeper and more accurate situational awareness of hazard events. We demonstrate how AHOM utilizes Twitter streams and bi-directional information exchange with social media users for enhanced hazard observation.

Modern Early Warning Systems (EWS) rely on scientific methods to analyse a variety of Earth Observation (EO) and ancillary data provided by multiple and heterogeneous data sources for the prediction and monitoring of hazard events. Furthermore, through social media, the general public can also contribute to the monitoring by reporting warning signs related to hazardous events. However, the warning signs reported by people require additional processing to verify the possibility of the occurrence of hazards. Such processing requires potential data sources to be discovered and accessed. However, the complexity and high variety of these data sources makes this particularly challenging. Moreover, sophisticated domain knowledge of natural hazards and risk management are also required to enable dynamic and timely decision making about serious hazards. In this paper we propose a data integration and analytics system which allows social media users to contribute to hazard monitoring and supports decision making for its prediction. We prototype the system using landslides as an example hazard. Essentially, the system consists of background knowledge about landslides as well as information about data sources to facilitate the process of data integration and analysis. The system also consists of an interactive agent that allows social media users to report their observations. Using the knowledge modelled within the system, the agent can raise an alert about a potential occurrence of landslides and perform new processes using the data sources suggested by the knowledge base to verify the event.