Spatial transcriptomics technologies have enabled comprehensive measurements of gene expression profiles while retaining spatial information, with most platforms also providing matched pathology images. However, noise resulting from low RNA capture efficiency and experimental steps needed to keep spatial information may corrupt the biological signals and obstruct analyses. Here, we develop a latent diffusion model DiffuST to denoise spatial transcriptomics. DiffuST employs a graph autoencoder and a pre-trained model to extract different-scale features from spatial information and pathology images. Then, a latent diffusion model is leveraged to map different scales of features to the same space for denoising. The evaluation based on various spatial transcriptomics datasets showed the superiority of DiffuST over existing denoising methods. Furthermore, the results demonstrated that DiffuST can enhance downstream analysis of spatial transcriptomics and yield significant biological insights.

Ensuring the safety of autonomous driving systems (ADS) requires stress testing under diverse and complex environmental conditions. Fuzzing has emerged as a promising approach, but existing methods often neglect the interplay between map complexity and multi-object mutations, resulting in limited coverage of safety-critical behaviors. To address this gap, we propose a map-aware reinforcement learning fuzzing (MARFT) approach that adaptively tailors fuzzing actions to spatial map features while jointly mutating multiple scenario parameters. Reinforcement learning, guided by undesirable behaviors and trajectory coverage, drives the selection of fuzzing actions to maximize exposure of safety-critical events through a closed-loop cycle of “scenario state perception → fuzzing action decision → driving reward feedback”. Experimental results demonstrate that our adaptive strategy significantly outperforms non-adaptive baselines in ADS failure detection. By combining Q-learning with trajectory coverage and behavior-driven seed selection, MARFT provides a scalable and adaptive solution for ADS testing, advancing the detection of critical behavioral patterns in complex driving environments.

Ensuring the safety of autonomous driving systems (ADSs) is essential, which requires effective testing methods to enhance system robustness. Fuzz testing (FT) is a widely used technique for uncovering software faults by generating test cases that trigger unexpected system behaviors. However, traditional FT in ADS suffers from significant limitations, including inefficient seed selection, low test case relevance, and inadequate exploration of diverse failure‐inducing driving scenarios. Random fuzzing often yields redundant or ineffective cases, limiting the detection of safety‐critical issues. To address these challenges, we propose ReinSeed, a reinforcement FT (RFT) framework that integrates three key phases: prefuzzing seed optimization, reinforcement learning (RL)–based scenario generation, and postfuzzing seed prioritization. We introduce a scenario complexity index to prioritize initial seeds before fuzzing. During fuzzing, we model the process as a Markov decision process (MDP) and apply Q ‐learning to generate scenarios with effective fuzzing action variations guided by driving behaviors, including undesired behaviors and trajectory coverage. To further improve testing effectiveness, we present a postfuzzing prioritization strategy that ranks fuzzed scenarios based on risk energy by incorporating control constraint violation analysis, safety‐critical events, and risk‐driven trajectory. Experimental results demonstrate that the unified framework—ReinSeed—significantly improves the detection of undesired behaviors, outperforming baseline methods across maps of varying complexity. Furthermore, the multiphase seed optimization showcases distinct contributions of scenario complexity, behavior‐guided fuzzing, and risk energy in enhancing both the efficiency and effectiveness of discovering critical behaviors in ADS.

Federated learning-based mobile crowdsourcing (F-MCS) leverages crowdsourcing for large-scale data perception, but it faces challenges from privacy concerns and network instability problems. Hence, privacy-protecting F-MCS schemes have been proposed to address these issues by aggregating local models on a trusted central server or a trusted third party (TTP). However, these schemes are still vulnerable to single points of failure and other malicious attacks, making them impractical. Moreover, due to the instability of the communication network, workers in the F-MCS scheme may drop out of the task, which oversees the entire model aggregation. In order to tackle the obstacles above, we design an aggregation method combined with Shamir secret sharing that comes with secure aggregation of global models without relying on a TTP. In addition, to enhance the robustness and adaptability of the scheme, we handle worker disconnection and new user joining to maintain protocol continuity and data integrity, thus tolerating dropouts and dynamic participation. We have conducted a thorough analysis of the scheme’s security, which shows that it can effectively protect user data privacy. Furthermore, our experimental results demonstrate that the proposed scheme performs well in model accuracy and is comparable to the system performance in the nondropout case.

In recent years, neural networks have been widely applied. However, adversarial attacks pose challenges to the secure deployment of neural networks. Adversarial training is one of the effective methods to train robust neural networks to resist such attacks. To address the high computational cost of adversarial training, we propose the masked adversarial contrastive distillation algorithm based on enhanced masked autoencoder (MAE-MACD). Unlike conventional approaches that demand the generation of adversarial samples in every iteration, MAE-MACD streamlines the process by requiring adversarial samples to be generated just once. First, MAE-MACD incorporates a feature learning module within the masked autoencoder, enabling the neural network to deduce global features from local ones. Second, it extracts the encoder and feature learning module from the masked autoencoder, utilizing them as the teacher model. In MACD, the knowledge distillation step involves training the model with different occlusion sizes and occlusion ratios, while the contrastive learning phase employs the same occlusion size but varying occlusion ratios to train the model. Finally, it fine-tunes the classification head using label information to ensure robust recognition performance. The experimental results demonstrate a significant improvement in neural network adversarial robustness achieved by MAE-MACD across CIFAR-10, CIFAR-100, and Tiny ImageNet, along with a reduced need for frequent adversarial sample generation.

As the demand for smart city services grows, unmanned aerial vehicle (UAV) swarm have achieved tremendous success in industries such as traffic management, logistics transportation, and road inspection. Despite their promising potential, a critical gap exists in the domain of drone swarm mission planning-a lack of a universal task planning method that can effectively address the complexities of diverse mission scenarios. To address this challenge, this paper introduces a novel scenario model-driven task planning method for UAV swarm. This method leverages scenario models as input, enabling the parsing of scenario tasks, UAV swarm resources, and scenario constraints. It subsequently facilitates multi-constraint task allocation through auction mechanisms and path planning via reinforcement learning. Through simulation experiments conducted in scenarios such as highway inspection and campus logistics, we validate the efficacy and versatility of the proposed method across different contexts.

An unmanned aerial vehicle (UAV) swarm is a cluster system composed of multiple UAVs and is widely used in military and civilian fields. The UAV swarm has a large number of resources, complex functions, space-time constraints, and task-driven characteristics. However, existing UAV swarm task description methods are usually limited to a specific task and cannot adapt to detailed descriptions of dynamic and complex application scenarios. To this end, we propose a UAV swarm application scenario model based on meta-level theory. Specifically, we abstract three types of meta models from UAV application scenarios: mission meta-model, resource meta-model, and constraint meta-model. Based on this model, we design and implement a UAV swarm application scenario modeling language (ASML) to support the formal description and analysis of the model. Furthermore, we define the conversion rules from ASML to timed automata. We model a logistics handling application scenario and use the model checking tool UPPAAL to verify the correctness of the scenario.

Researchers have been developing AI-based mod-eling assistants to help software modelers efficiently construct models. However, there are a number of issues with the current modeling assistants, including poor recommendation accuracy, limited support for diverse model types, and scalability issues. These problems stem from their attempt to utilize a single learning module to comprehensively extract multi-modal features from software models, such as semantic meanings of terms and model structures. Our key insight is that the utilization of modular deep learning architecture, allowing these features to be learned separately, by specifically tailored neural modules, and then be fused into one single vector. The fused vectors produced by these two learning stages significantly enhance recommendation accuracy. To adapt model formats for modular learning's input, we introduce a novel model representation, labeled graph, which offers two advantages: 1) able to segregate diverse features types, enabling modular learning; 2) adaptable to various types of software models. Building on these insights, we developed ReCo, a learning-based recommendation system for suggesting connections in models. ReCo employs several neural modules for extracting both the semantics of elements and topology of models, then computing the scores of potential connections. Our experimental result shows that for model types that are already supported, ReCo achieves more than 2X improvement in success rate and FRanks, compared to the state-of-the-art modeling assistants. Furthermore, ReCo also extends its support to previously unsupported models like UML usecase and activity models.

Summary Software vulnerability detection is an important problem in software security. In recent years, deep learning offers a novel approach for source code vulnerability detection. Due to the similarities between programming languages and natural languages, many natural language processing techniques have been applied to vulnerability detection tasks. However, specific problems within vulnerability detection tasks, such as buffer overflow, involve numerical reasoning. For these problems, the model needs to not only consider long dependencies and multiple relationships between statements of code but also capture the magnitude property of numerical literals in the program through high‐quality number embeddings. Therefore, we propose VDTransformer, a Transformer‐based method that improves source code embedding by integrating word and number embeddings. Furthermore, we employ Transformer encoders to construct a hierarchical neural network that extracts semantic features from the code and enables line‐level vulnerability detection. To evaluate the effectiveness of the method, we construct a dataset named OverflowGen based on templates for buffer overflow. Experimental comparisons on OverflowGen with a well‐known static vulnerability detector and two state‐of‐the‐art deep learning‐based methods confirm the effectiveness of VDTransformer and the importance of high‐quality number embeddings in vulnerability detection tasks involving numerical features.

Cyber-physical systems (CPSs) are becoming increasingly important in safety-critical systems. Particular risk analysis (PRA) is an essential step in the safety assessment process to guarantee the quality of a system in the early phase of system development. Human factors like the physical environment are the most important part of particular risk assessment. Therefore, it is necessary to analyze the safety of the system considering human factor and physical factor. In this paper, we propose a new particular risk model (PRM) to improve the modeling ability of the Architecture Analysis and Design Language (AADL). An architecture-based PRA method is presented to support safety assessment for the AADL model of a cyber-physical system. To simulate the PRM with the proposed PRA method, model transformation from PRM to a deterministic and stochastic Petri net model is implemented. Finally, a case study on the power grid system of CPS is modeled and analyzed using the proposed method.

Software Testing is an approach to ensuring the quality of software systems. Testing of safety-critical systems often requires conformance to certain code coverage criteria, including for example, in aviation, Modified Condition/Decision Coverage (MC/DC). In some situations, however, access to the actual code may be restricted with black Box approaches, and testers may only be able to use models of the system, such as those in MATLAB's Simulink. Without access to the code, exact code coverage measurement may not be possible. This paper presents a method of identifying and using the Simulink model's constraints to generate test cases which can achieve high coverage of the actual source code. A case study confirming the relationship between the model's coverage and the code coverage is also presented.

Quantitative verification is an effective technique for analyzing quantitative aspects of a safety critical system's design, and safety analysis is a significant aspect of safety critical system. However, they are often conducted separately. In this paper, we propose a new methodology, QaSten, fastens quantitative verification to safety analysis for Architecture Analysis and Design Language (AADL) model (including error model). QaSten formalizes a set of rigorous transformation rules that transform AADL model to PRISM model using formal method. In addition, QaSten can generate two safety property formulas automatically to check against the PRISM model for each hazardous state. Therefore, the occurrence probability of hazardous states can be calculated, which can help system designers understand the impact of parameters in the model. Furthermore, combining the probability and the severity of potential consequence of a hazardous state, QaSten determines the hazard risk acceptance level that can help engineers to identify critical hazard and modify or redesign architecture model to control it in an acceptable level. Two case studies, based on the Gas Leakage Alarm systems, are utilized to demonstrate QaSten's feasibility and effectiveness.

Safety analysis is a significant aspect of safety critical embedded systems. In this paper, an architecture-based hazard analysis method is presented to support safety assessment for Architecture Analysis and Design Language (AADL) model of embedded systems during early development phases. For further improving the hazard analytical ability of AADL, Hazard Model Annex is created. In order to improve the quality of system and the software development process, a safety model can be established by extending AADL model with error model and hazard model to specify fault behavior and hazard behavior of system. Hazard factor can be identified in safety model through hazard analysis. Additionally, conversion rules and formal methods are formulated to transform safety model into Deterministic Stochastic Petri Net (DSPN) for quantitative analysis using an existing tool. Finally, a safety analysis table is generated for overall evaluation of hazards, including hazard risk acceptance level, to help engineers to eliminate or control component hazards in an acceptance level. A small case study, based on fire alarm system, is utilized to demonstrate the feasibility of hazard analysis method for AADL model.

As the limitations of energy consumption for real-time embedded systems more strict, it has been difficult to ignore the context switch overhead for Fixed-Priority task with Preemption scheduling (FPP) in multitasking environment. This paper presents a Reducing Context Switches scheduling (RCSS) based on preemption thresholds scheduling for real-time embedded system to decrease system energy consumption. The WCRT model is improved based on considering context switch overhead. In addition, the tasks energy consumption is analyzed. The experimental results show that RCSS can reduce context switches about 9.051‰ and decrease energy consumption about 6.129‰ for given tasks compared to FPP.

Spectra-based fault localization (SFL) is an automatic fault-localization technique which has received a lot of attention due to its simplicity and effectiveness. SFL uses ranking metric (RM) to rank the risk of fault existence in each program entity after dynamically collecting the necessary information. The evaluation of RMs for SFL has recently become a research focus. To evaluate the average performance of RMs for SFL with different single-fault types, an If-While-If (IWI) model-based approach is presented in this paper. Firstly, through investigating rankings of statements in the IWI model, this paper takes an optimal RM known as an example to analyze its localization effectiveness for five types of single-fault. Secondly, a generic hierarchical method is given in the IWI model to precisely calculate the average performance of RMs. Two experiments, that calculate the average performance of the optimal RM on the IWI model and actual programs, are conducted with five single-fault types. The experimental results agree with theoretical analyses. It is found that the average performance of the optimal RM is related to the number of test cases and the number of program cycles, and the fault type. The IWI model could function as large programs to effectively evaluate RMs for different fault types.

Cyber-Physical Systems (CPS) tightly integrate cyber and physical components and transcend discrete and continuous domains. It is greatly desired that the physical components being controlled and the software implementation of control algorithms can be verified together. We present an efficient approach to reachability analysis of Hybrid Automata Pushdown System (HAPS) models for cyber/physical co-verification of CPS. We have realized this approach and applied it to real-world control systems. The evaluation has shown that HAPS is an effective model for co-verification of CPS and our approach has major potential in verifying system-level properties of CPS, therefore improving the reliability of CPS.

Software reliability models are used to estimation and prediction of software reliability. Existing models either use black-box approach that based on test data during software test phase or white-box approach that based on software architecture and individual component reliability, which is more suited to assess the reliability of modern software system. However, most of the reliability models based on architecture assumed that a failure occurring within one component will not cause any other component to fail, which is inconsistent with the facts. This paper introduces a reliability model and a reliability analysis technique for architecture-based reliability evaluation. Our approach extend existing reliability model by considering fault propagation. We believe that this model can be used to effectively improve software quality.