Today, the railway industry is experiencing a profound digital transformation driven by advances in ICT technologies such as 5G, cloud computing, and artificial intelligence. In particular, the digitalization of rail control systems has significant potential to reduce system complexity and operating and maintenance costs while improving safety and efficiency. However, in addition to these significant benefits, security has emerged as a critical concern that could no longer be overlooked. Even a minor breach in any digitalized component can propagate through interconnected subsystems and eventually lead to severe safety incidents in the physical world. To holistically address both safety and security in digitalized rail control systems, we propose an adaptive cyber-resilience paradigm by analyzing the characteristics, risks, and requirements of the system. As safety has been extensively studied in the rail domain, this paper focuses primarily on cybersecurity and its interaction and integration with safety. By modeling the railway and its control system as a specialized form of Cyber-Physical System (CPS), we elucidate cybersecurity requirements and introduce a comprehensive risk-analysis methodology to identify the key challenges to achieve cyber resilience. We further define the cyber-resilience paradigm and conduct an in-depth examination of its four essential components: theoretic models, enabling techniques, standards, and practical implementations. Finally, we develop a MAPE-K-based approach and implement this paradigm in the Chinese Train Control System (CTCS), thus demonstrating its feasibility and practical applicability.

Microservice architectures in modern software engineering generate vast quantities of heterogeneous metrics, making fault diagnosis notoriously difficult. Conventional root cause analysis (RCA) methods often struggle with high-dimensional, diverse data where only a small subset of metrics may truly drive the observed failures. In this paper, we propose CAVIAR (Causality-based Analysis via VAE and ICA for Anomaly Root-cause), a two-phase framework for interpretable RCA in large-scale microservice systems. First, we train a variational autoencoder (VAE) enhanced with Independent Component Analysis (ICA) principles to learn a lowdimensional, disentangled representation of normal microservice operation. By enforcing independence among latent variables, we discover semantically coherent factors, such as specific service loads or network-level conditions. Second, when a fault occurs, we treat anomalies as external interventions on some latent factor and optimize an interventional matrix to identify the culprit dimension. This factor is then mapped back to the original metrics for actionable diagnostics. Empirical results on both industrial-scale and publicly available datasets show that CAVIAR outperforms baseline approaches in accurately localizing root causes-even at high dimensionalities where graph-based or homogeneous-metric methods fail. A detailed case study on a microservice platform demonstrates that our approach naturally uncovers the hidden propagation paths of failures, offering a clear explanation of how a single overloaded service can trigger system-wide latency breaches. We conclude that ICA-based VAEs provide a scalable, interpretable alternative for fault diagnosis, bridging the gap between deep generative modeling and practical software engineering needs.

Acute kidney injury (AKI) represents a life-threatening condition among hospitalized patients, where early prediction enables prevention. Despite advances in existing models, clinical implementation remains hindered by excessive false positive rates (70%-94%) and lack of actionable clinical insights. We conduct a multi-center retrospective cohort study and develop a two-model large language model framework: AKI-PM (Prediction Model) for predicting AKI occurrence within 24 hours and AKI-RAM (Risk Attribution Model) for providing explainable risk attribution. Using a cohort of 140,637 hospital admissions across four geographically diverse Chinese hospitals, we demonstrate that AKI-PM achieves high predictive performance in internal validation (area under curve 0.95, positive predictive value 0.68) and maintains robust generalizability across external sites after few-shot (area under curve 0.92-0.96, positive predictive value 0.69-0.74). Crucially, AKI-RAM provides structured, clinically actionable risk explanations by distinguishing modifiable from non-modifiable factors and offering tailored recommendations. In a clinical evaluation of 200 cases from four independent hospitals by six nephrologists, AKI-RAM receives high scores across eight dimensions (Likert scale: 4.18-4.88) with moderate to good inter-rater reliability (intraclass correlation coefficients: 0.680-0.803). This integrated framework addresses critical limitations in AI-driven clinical prediction by combining accuracy with interpretability, offering a scalable solution for early AKI prevention in diverse healthcare settings.

The Segment Anything Model (SAM) has been introduced for universal segmentation in natural images but struggles with medical image segmentation due to a significant domain gap. While various medical SAM variants have been developed to adapt SAM to medical datasets, we observe that they still require a large amount of labeled data to perform effectively in digital pathology semantic segmentation, where annotations are particularly costly and time-consuming. To address this challenge, we propose SAM-ALE, a novel approach that enhances SAM for Low-shot digital pathology semantic segmentation via an Auxiliary Lightweight Encoder. Our approach leverages a small amount of labeled data and a large pool of unlabeled data to train a task-specific lightweight encoder, which captures pathology-specific visual features. The encoded features are then integrated with SAM’s embeddings to improve segmentation performance with minimal labeled supervision. Experiments on the CRAG and MoNuSeg datasets validate our approach. For example, with just 2% labeled data, our method enables SAM to achieve performance (86.52% Dice) comparable to its fully fine-tuned counterpart (87.37% Dice) on CRAG, utilizing a lightweight encoder with only 7.7M parameters. Additionally, our approach allows seamless integration with various medical SAM variants. The code and data are released at https://github.com/qianyuli123/SAM-ALE.

With the explosive growth of global data volume, the usage of hard disk drives (HDDs) is also increasing rapidly. Consequently, the number of failed disks is continuously rising, which can affect storage service quality and even lead to data loss when failures occur.In recent years, the active fault-tolerant technology, which collects hard disks' Self Monitoring Analysis and Reporting Technology (SMART) data-set, predicts hard disk failure by machine learning model, and repairs near-failure disks' data to health disks in advance, has become a common research hotspot in both academia and industry. Aiming at the existing problems such as interference characteristics, inaccurate failure time prediction, competition of system resources between data migration and front service, this paper researches the two-stage prediction model and data migration strategy based on hard disk failure time, including the two-stage hard disk information feature selection method, the two-stage prediction method of hard disk failure time, and the data migration elastic system resource allocation strategy. Feature selection is performed by combining embedding methods with visualization, and the importance of the selected features is evaluated using a random forest model. Based on the feature importance, further refinement is carried out to obtain the final feature set. Before predicting the failure time of hard drives, XGBoost is first used in a voting manner to identify drives predicted to be faulty. Then, a trained Bidirectional Long ShortTerm Memory network (Bidirectional LSTM) enhanced with a self-attention mechanism is employed to predict the exact failure time.Experimental results show that on the Backblaze dataset, the model achieves a mean absolute error of 1.24 when predicting failure times. The recall rate for predicting failures within 7 days reaches 98.79 %, the error rate is 0.30 %, the F1 score is 99.24 %, and the precision is 99.69 %. The elastic system resource allocation strategy for data migration improves business IOPS by 47.19 % and reduces latency by 38.68 %.

File system caching is crucial for enhancing I/O performance; yet traditional caching strategies exhibit low efficiency when handling complex, non-sequential, and data-intensive workloads characterized by concurrent multi-user access, as typified by AI and database applications. Existing learning-based prefetching research faces a dual challenge of mismatched prediction granularity and the absence of semantic information. This paper proposes GFPP (Graph-based File Prefetcher), a deep graph neural network framework operating at the file system level. GFPP innovatively employs a parallel spatio-temporal decoupled architecture: a Graph Neural Network (GNN) branch extracts microscopic spatial-topological structures from dynamic I/O interaction graphs, while a multi-scale temporal Convolutional Neural Network (CNN) branch concurrently captures macroscopic long-range sequential patterns; simultaneously integrating a Dynamic-K optimization mechanism. In large-scale real-world cloud platform I/O workloads, GFPP’s cache hit rate is significantly superior to advanced baseline models such as Transformer, GAT, and SGDP. Furthermore, in zero-shot generalization tests on Web application and database workloads, the strong adaptability and robustness of GFPP under unseen workloads are demonstrated. This research reveals the core value of fine-grained spatio-temporal collaborative modeling at the file system level, providing an effective framework and a novel perspective for building next-generation high-performance intelligent caching systems.

This article proposes PowerCause, a method for phase unbalance positioning and active regulation in smart distribution networks. PowerCause can automatically detect anomaly intervals, locate the source of phase unbalance using Granger causality test, back-search, and generate a list of potential root cause buses. It takes regulation measures for specific buses to alleviate the impact of the unbalance. This implements a closed-loop solution that handles the entire process from the occurrence of phase unbalance to its positioning, and finally, to active regulation. This study builds a closed-loop simulation environment based on the open distribution system simulator (OpenDSS) to enable autonomous and controllable unbalance injection, collect multidimensional bus metrics, including voltage and phase angle. The environment also includes causal analysis and active regulation modules for method verification. The proposed method demonstrates high accuracy in root cause location, efficient performance, and robustness against environmental influences such as measurement noise and data loss errors.

Recent advances in medical multi-modal models focus on specialized image analysis like dermatology, pathology, or radiology. However, they do not fully capture the complexity of real-world clinical diagnostics, which involve heterogeneous inputs and require ongoing contextual understanding during patient-physician interactions. To bridge this gap, we introduce PulseMind, a new family of multi-modal diagnostic models that integrates a systematically curated dataset, a comprehensive evaluation benchmark, and a tailored training framework. Specifically, we first construct a diagnostic dataset, MediScope, which comprises 98,000 real-world multi-turn consultations and 601,500 medical images, spanning over 10 major clinical departments and more than 200 sub-specialties. Then, to better reflect the requirements of real-world clinical diagnosis, we develop the PulseMind Benchmark, a multi-turn diagnostic consultation benchmark with a four-dimensional evaluation protocol comprising proactiveness, accuracy, usefulness, and language quality. Finally, we design a training framework tailored for multi-modal clinical diagnostics, centered around a core component named Comparison-based Reinforcement Policy Optimization (CRPO). Compared to absolute score rewards, CRPO uses relative preference signals from multi-dimensional com-parisons to provide stable and human-aligned training guidance. Extensive experiments demonstrate that PulseMind achieves competitive performance on both the diagnostic consultation benchmark and public medical benchmarks.

Recent works explore how real and synthetic data contribute to Vision-Language-Action (VLA) models' generalization. While current VLA models have shown the strong effectiveness of large-scale real-robot pre-training, synthetic data has not previously demonstrated comparable capability at scale. This paper provides the first evidence that synthetic data alone can match the performance of the strongest π-dataset in pre-training a VLA model, revealing the substantial value of large-scale simulation. The resulting model also exhibits surprisingly zero-shot sim-to-real transfer on several challenging tasks. Our synthetic dataset, InternData-A1, contains over 630k trajectories and 7,433 hours across 4 embodiments, 18 skills, 70 tasks, and 227 scenes, covering rigid, articulated, deformable, and fluid-object manipulation. It is generated through a highly autonomous, fully decoupled, and compositional simulation pipeline that enables long-horizon skill composition, flexible task assembly, and heterogeneous embodiments with minimal manual tuning. Using the same architecture as π_0, we pre-train a model entirely on InternData-A1 and find that it matches the official π_0 across 49 simulation tasks, 5 real-world tasks, and 4 long-horizon dexterous tasks. We release the dataset and will open-source the generation pipeline to broaden access to large-scale robotic data and to lower the barrier to scalable data creation for embodied AI research.

Shuffling has been shown to amplify differential privacy guarantees, enabling a more favorable privacy-utility trade-off. To characterize and compute this amplification, two fundamental analytical frameworks have been proposed: the privacy blanket by Balle et al. (CRYPTO 2019) and the clone paradigm–including both the standard and stronger variants–by Feldman et al. (FOCS 2021, SODA 2023). These frameworks share a common foundation: decomposing local randomizers into structured components for analysis. In this work, we introduce a unified analytical framework–the general clone paradigm–which subsumes all possible decompositions, with the clone and blanket decompositions arising as special cases. Within this framework, we identify the optimal decomposition, which is precisely the one used by the privacy blanket. Moreover, we develop a simple and efficient algorithm based on the Fast Fourier Transform (FFT) to compute optimal privacy amplification bounds. Experimental results show that our computed upper bounds nearly match the empirical lower bounds, demonstrating the tightness of our method. Building on this method, we also derive optimal amplification bounds for both joint and parallel compositions of LDP mechanisms in the shuffle model.

The shuffle model, which anonymizes data by randomly permuting user messages, has been widely adopted in both cryptography and differential privacy. In this work, we present the first systematic study of the Bayesian advantage in re-identifying a user's message under the shuffle model. We begin with a basic setting: one sample is drawn from a distribution P, and n - 1 samples are drawn from a distribution Q, after which all n samples are randomly shuffled. We define β_n(P, Q) as the success probability of a Bayes-optimal adversary in identifying the sample from P, and define the additive and multiplicative Bayesian advantages as 𝖠𝖽𝗏_n^+(P, Q) = β_n(P,Q) - 1/n and 𝖠𝖽𝗏_n^×(P, Q) = n · β_n(P,Q), respectively. We derive exact analytical expressions and asymptotic characterizations of β_n(P, Q), along with evaluations in several representative scenarios. Furthermore, we establish (nearly) tight mutual bounds between the additive Bayesian advantage and the total variation distance. Finally, we extend our analysis beyond the basic setting and present, for the first time, an upper bound on the success probability of Bayesian attacks in shuffle differential privacy. Specifically, when the outputs of n users–each processed through an ε-differentially private local randomizer–are shuffled, the probability that an attacker successfully re-identifies any target user's message is at most e^ε/n.

The shuffle model offers significant privacy amplification over local differential privacy (LDP), enabling improved privacy-utility trade-offs. To analyze and quantify this amplification effect, two primary frameworks have been proposed: the \textit{privacy blanket} (Balle et al., CRYPTO 2019) and the \textit{clone paradigm}, which includes both the \textit{standard clone} and \textit{stronger clone} (Feldman et al., FOCS 2021; SODA 2023). All of these approaches are grounded in decomposing the behavior of local randomizers. In this work, we present a unified perspective--termed the \textit{general clone paradigm}--that captures all decomposition-based analyses. We identify the optimal decomposition within this framework and design a simple yet efficient algorithm based on the Fast Fourier Transform (FFT) to compute tight privacy amplification bounds. Empirical results show that our computed upper bounds nearly match the corresponding lower bounds, demonstrating the accuracy and tightness of our method. Furthermore, we apply our algorithm to derive optimal privacy amplification bounds for both joint composition and parallel composition of LDP mechanisms in the shuffle model.

Early prediction of acute kidney injury (AKI) may provide a crucial opportunity for AKI prevention. To date, no prediction model targeting AKI among general hospitalized patients in developing countries has been published. Here we show a simple, real-time, interpretable AKI prediction model for general hospitalized patients developed from a large tertiary hospital in China, which has been validated across five independent, geographically distinct, different tiered hospitals. The model containing 20 readily available variables demonstrates consistent, high levels of predictive discrimination in validation cohort, with AUCs for serum creatinine-based AKI and severe AKI within 48 h ranging from 0.74–0.85 and 0.83–0.90 for transported models and from 0.81–0.90 and 0.88–0.95 for refitted models, respectively. With optimal probability cutoffs, the refitted model could predict AKI at a median of 72 (24–198) hours in advance in internal validation, and 54–90 h in advance in external validation. Broad application of the model in the future may provide an effective, convenient and cost-effective approach for AKI prevention. Early prediction of Acute kidney injury (AKI) may be crucial for AKI prevention. Here the authors present a simple, real-time, interpretable, AKI prediction model for hospitalized patients, based on routinely collected electronic health records data.

Uncovering cause-effect relationships from observational time series is fundamental to understanding complex systems. While many methods infer static causal graphs, real-world systems often exhibit *dynamic causality*—where relationships evolve over time. Accurately capturing these temporal dynamics requires time-resolved causal graphs. We propose UnCLe, a novel deep learning method for scalable dynamic causal discovery. UnCLe employs a pair of Uncoupler and Recoupler networks to disentangle input time series into semantic representations and learns inter-variable dependencies via auto-regressive Dependency Matrices. It estimates dynamic causal influences by analyzing datapoint-wise prediction errors induced by temporal perturbations. Extensive experiments demonstrate that UnCLe not only outperforms state-of-the-art baselines on static causal discovery benchmarks but, more importantly, exhibits a unique capability to accurately capture and represent evolving temporal causality in both synthetic and real-world dynamic systems (e.g., human motion). UnCLe offers a promising approach for revealing the underlying, time-varying mechanisms of complex phenomena.

Password vaults are used to manage multiple account passwords, encrypted with a master password. However, cipher-text stored on synchronization servers is vulnerable to leakage and offline guessing attacks, potentially compromising all accounts. Honey password vaults address this by generating decoy vaults for incorrect master passwords, making offline guessing infeasible and requiring online verification. Existing studies on honey vaults rely on a small dataset of only 276 vaults for model training and security evaluation, limiting their conclusions. More importantly, existing evaluations focus solely on the distinguishability between real and decoy vaults, overlooking practical security: How many accounts could be cracked via online guessing? In this paper, we construct a large dataset of millions of vaults by aggregating numerous leaked password datasets. With the dataset, we employ advanced machine learning techniques for both decoy generation and identification. We show that various text classification algorithms, especially pre-trained models, significantly outperform existing attacks with distinguishing accuracy of 95.79%-83.75%. Further, we introduce a Transformer model that generates more plausible decoy vaults, no attacks achieve accuracy more than 64.35%. We further assess the practical security of honey vaults against online guessing. Our new model achieves the best performance, only 0.51 accounts is cracked on average with 1,000 online attempts. By applying two simple measures, we enhance the scheme to a practical level: 1) using honey accounts for leakage detection, and 2) avoiding the encryption of passwords for websites with unlimited login attempts. These improvements reduce the cracked number to 0.11. We also offer new insights, such as that even a poor model can achieve notable practical security by using our measures.

In the Internet of Everything (IoE) era, various sensors configured at the Internet of Things (IoT) edge (Edge) collect massive amounts of data daily. Hard drives are the most failure-prone component in Edge, leading to service unavailability and the permanent loss of sensor-collected data. To solve this problem, we propose Cescpra, a cloud-edge-sensor collaborative proactive reliability assurance technology. Sensors collect data and store it at the Edge using erasure coding, with data and parity blocks stored separately on the Edge and in the cloud, periodically updated asynchronously. To enhance reliability, we propose a two-level model strategy for hard drive failure prediction, with lightweight models on each device and a full-weight model in the cloud. Samples predicted as suspected faults by the lightweight model will be sent to the cloud for final accurate prediction. Upon confirming imminent failure, proactive data recovery through migration and rebuild is initiated. In an experimental setup with a three-node cloud server, 42 IoT edges, and 210 sensors, our hard drive failure prediction model achieved a true-positive rate (TPR) of 99%, an area under the curve (AUC) score of 0.999, and a false-positive rate (FPR) below 0.01%, minimizing unnecessary data recovery. The multidisk joint proactive recovery method reduced fault tolerance time by up to 64.7% and 62.5% compared to individual migration or rebuild methods. This cloud-edge-sensor collaborative architecture leverages Edge's computational and storage capabilities, significantly enhancing the reliability of sensor data.

In cloud storage services, burst I/O workloads from data analytics and artificial intelligence/machine learning (AI/ML) applications present significant challenges to Quality of Service (QoS) management. Existing scheduling models like dmClock ensure fair and stable I/O bandwidth allocation in typical scenarios. However, they falter under frequent burst traffic, leading to lower resource utilization and higher task latency. To address this, we propose BAQoS, a Burst I/O Aware Quality of Service Optimization for Cloud Storage Service. BAQoS employs refined request classification, a burst-aware hierarchical scheduling algorithm, and a high-performance scheduler architecture (HPSA). These features enable dynamic resource allocation and efficient scheduling for both burst and regular requests. Experiments show that BAQoS markedly enhances performance under burst workloads, accelerating burst request processing by up to 7.09 and improving overall system performance by 48.86%. Furthermore, BAQoS ensures superior performance for non-burst users, achieving a 51.15% performance boost for high-priority users, a 5.34% increase in system throughput, and an over 50% reduction in IOPS standard deviation among same-priority users.

As password breaches increase, users frequently face multiple password leaks. Because of memory limitations, users often reuse or slightly modify their passwords across different accounts. This behavior makes targeted password-guessing attacks, which leverage leaked passwords, a serious security threat. While previous studies have primarily focused on single-leak scenarios, where the attacker possesses only one leaked password, multi-leak scenarios remain insufficiently explored. In this work, we propose Pass2Pass-T, a model designed to capture similarities across multiple passwords from the same user. Pass2Pass-T leverages Transformers to predict a target password from multiple leaked passwords of the same user. Additionally, we are the first to empirically evaluate the multi-leak attacks on real-world password datasets. In multi-leak scenarios, Pass2Pass-T enhances the Transformer with input compression and segmented positional encoding, tailored to the distinct characteristics of password sequences unlike natural language texts. It also utilizes transfer learning to effectively model patterns across multiple passwords. With five leaked passwords, our model compromises 4.87 × improvement over state-of-the-art strategies that process each leaked password individually. In single-leak scenarios, Pass2Pass-T matches existing models at 1,000 guesses and achieves a 4.46 10^7 guesses.

Variational Autoencoder (VAE) has been used as password generative model for trawling attack in multiple works. Its sample distribution can be easily changed by controling the mean and variance of the prior distribution, which makes it natively suitable for dynamic attack scenario. Combining transformer blocks with VAE can achieve better performance since attention mechanisms can handle sequence data better. But such design is unstable for password generation tasks. The encoder-decoder model tends to degrade into decoder-only model due to the KL vanishing problem, making it hard to train. To handle this problem, we performed an in-depth analysis and proposed a new transformer-based VAE model specifically designed for password generation. It out-performs former encoder-decoder generative model by 4%-15% in cracking rate. Moreover, we make an improvement to dynamic attack by using a 3-period strategy, with which our method becomes competitive with probabilistic ordered attack models such as PCFG [11] and FLA [8].