Downhill curved segments of mountainous freeways are high-risk locations due to the combined effects of longitudinal potential energy and lateral centrifugal forces, and speed limit optimization is a key measure for mitigating crash risk on such segments. However, conventional speed limit settings often rely on uniform standards and fail to adequately account for the combined influences of roadway geometry, slope, traffic conditions, and weather, resulting in heterogeneous safety outcomes across different road environments. This study employs a Causal Forest model with explicitly incorporated quarterly time variables, complemented by a Difference-in-Differences (DID) model, to evaluate the safety effectiveness of speed limit optimization on downhill curved segments of the Guidu Freeway. Empirical results indicate that speed limit measures significantly reduce crash risk, with the Causal Forest model estimating an average treatment effect (ATE) of -0.203 (95% confidence interval: -0.213 to -0.193), demonstrating high precision and robustness and outperforming the DID model in estimation stability. Heterogeneity analysis further reveals that the safety benefits are most pronounced on segments with relatively mild horizontal curvature, low curvature variability, and moderate downhill slopes, as well as under moderate traffic volumes (approximately 7,500-9,000 vehicles/day), while higher truck proportions weaken the effectiveness of the measures. In contrast, fog frequency exerts a relatively limited moderating effect, as treatment effects remain negative across different fog conditions. Overall, the findings confirm that speed limit optimization can substantially improve safety on downhill curved segments of mountainous freeways and highlight the importance of accounting for roadway geometry and traffic composition when designing targeted speed management strategies.

This study clarified the Tor network architecture and website fingerprinting security risks from side-channel leaks of packet timing and traffic bursts in encrypted traffic. It proposed a hierarchical classification framework for website fingerprinting defenses, and horizontally and vertically compared four defense categories: traffic feature obfuscation, attack model misdirection, anonymity set enhancement, and cross-layer collaborative optimization. It reviewed public evaluation frameworks and datasets, and constructed an evaluation system integrating defense effectiveness, system overhead and deployment characteristics for defense selection. Core challenges including concept drift and adversarial robustness are summarized, and frontier research directions such as adaptive defense and native Tor integration are outlined.

The Tor network’s anonymity is increasingly exploited for cybercrime, creating a demand for accurate traffic classification under strict few-shot constraints. While recent efforts like WF-Transformer demonstrate strong temporal modeling capabilities, they still require abundant labeled data and struggle to generalize under defense-induced distortions and open-world unknown traffic. To address these gaps, we propose STNet (SpatioTemporal Multi-scale Augmentation and fusion Network), an episode-based few-shot learning architecture for Tor traffic classification. Unlike simple module stacking, STNet adopts a modular decoupling design: (1) a Multi-Scale Spatiotemporal Feature Fusion (MSMF) module captures packet-level and flow-level patterns to resist obfuscation; (2) scenario-adaptive modules tackle domain shifts in closed-world settings and feature scarcity in open-world settings; and (3) a Hierarchical Layer Attention (HLA) mechanism dynamically fuses heterogeneous features from different deployment positions. Extensive experiments on real-world Tor traffic show that STNet consistently outperforms representative baselines including WF-Transformer. In closed-world settings, it limits the accuracy drop under WalkieTalkie obfuscation to 13.6 percentage points. In open-world 10-shot evaluation, it achieves 92.1% AU-COVR and 79.1% unknown-class F1-score, surpassing the best baseline by 4.9 and 6.0 percentage points, respectively. These results demonstrate the effectiveness of decoupling universal feature extraction from scenario-specific adaptation in few-shot Tor traffic analysis.

Unobserved heterogeneity is a key factor leading to biased estimates and unreliable crash predictions, undermining the effectiveness of safety countermeasures. Two common options to address it are separating crash types and leveraging advanced statistical models. While finite mixture models have recently demonstrated superiority in this regard, they fail to account for correlations between different crash types. This study thus integrates the two approaches by developing bivariate finite mixture (BFM) models while decomposing crashes into single‐vehicle (SV) and multivehicle (MV) crashes so as to simultaneously address the correlation feature and unobserved heterogeneity. Two variants of the BFM model are developed with SV and MV crash counts as the bivariate response variable. The first variant, BFM‐I‐2, measures the correlation using multivariate normally distributed random error terms, while the second, BFM‐II‐2, employs a joint probability density function. Both bivariate models categorize SV and MV crashes into two subgroups, respectively. Among these models, BFM‐I‐2 exhibits better performance than BFM‐II‐2; it not only verifies the positive correlations between SV and MV crashes but also highlights substantial differences in the key contributing factors across the two crash types. Traffic volume exerts opposing effects on the two subgroups of SV crashes, indicating that unobserved heterogeneity may explain the inconsistent findings regarding its influence in previous research. For MV crashes, the two subgroups share few common risk factors: those in Component 1 are predominantly associated with traffic flow conditions, whereas Component 2 is largely affected by road geometry. These findings can support developing tailored freeway safety interventions.

Malware remains a central tool in cyberattacks, and systematic research into adversarial attack techniques targeting malware is crucial in advancing detection and defense systems that can evolve over time. Although numerous review articles already exist in this area, there is still a lack of comprehensive exploration into emerging artificial intelligence technologies such as reinforcement learning from the attacker’s perspective. To address this gap, we propose a foundational reinforcement learning (RL)-based framework for adversarial malware generation and develop a systematic evaluation methodology to dissect the internal mechanisms of generative models across multiple key dimensions, including action space design, state space representation, and reward function construction. Drawing from a comprehensive review and synthesis of the existing literature, we identify several core findings. (1) The scale of the action space directly affects the model training efficiency. Meanwhile, factors such as the action diversity, operation determinism, execution order, and modification ratio indirectly influence the quality of the generated adversarial samples. (2) Comprehensive and sensitive state feature representations can compensate for the information loss caused by binary feedback from real-world detection engines, thereby enhancing both the effectiveness and stability of attacks. (3) A multi-dimensional reward signal effectively mitigates the policy fragility associated with single-metric rewards, improving the agent’s adaptability in complex environments. (4) While the current RL frameworks applied to malware generation exhibit diverse architectures, they share a common core: the modeling of discrete action spaces and continuous state spaces. In addition, this work explores future research directions in the area of adversarial malware generation and outlines the open challenges and critical issues faced by defenders in responding to such threats. Our goal is to provide both a theoretical foundation and practical guidance for building more robust and adaptive security detection mechanisms.

For attacker's intranet lateral movement stage of the cyber kill chain, current cyber defense strategies face the problems of lacking dynamism, being unable to make realtime decisions, and failure to take into account the dynamic changes of attacker's strategies. In order to solve the above problems, this paper proposes a new gridworld model to describe the lateral movement scenario. The multi-agent stochastic game is used to model attack and defense process, and the Nash Q-learning algorithm is used to solve the optimal cyber defense strategy. Through multiple simulation experiments, it turns out that the proposed model has good convergence, and the average defense success rate of the defender is more than 70%, which verifies the effectiveness of the model.

Adversarial training of lightweight models faces poor effectiveness problem due to the limited model size and the difficult optimization of loss with hard labels. Adversarial distillation is a potential solution to the problem, in which the knowledge from large adversarially pre-trained teachers is used to guide the lightweight models’ learning. However, adversarially pre-training teachers is computationally expensive due to the need for iterative gradient steps concerning the inputs. Additionally, the reliability of guidance from teachers diminishes as lightweight models become more robust. In this paper, we propose an adversarial distillation method called Sample-Adaptive Multi-teacher Dynamic Rectification Adversarial Distillation (SA-MDRAD). First, an adversarial distillation framework of distilling logits and features from the heterogeneous standard pre-trained teachers is developed to reduce pre-training expenses and improve knowledge diversity. Second, the knowledge of teachers is distilled into the lightweight model after sample-aware dynamic rectification and adaptive fusion based on teachers’ predictions to improve the reliability of knowledge. Experiments are conducted to evaluate the performance of the proposed method on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. The results demonstrate that our SA-MDRAD is more effective than existing adversarial distillation methods in enhancing the robustness of lightweight image classification models against various adversarial attacks.

In the face of an increasingly intricate network structure and a multitude of security threats, cyber deception defenders often employ deception assets to safeguard critical real assets. However, when it comes to the intranet lateral movement attackers in the cyber kill chain, the deployment of deception assets confronts the challenges of lack of dynamics, inability to make real-time decisions, and not considering the dynamic change of an attacker’s strategy. To address these issues, this study introduces a novel maze pathfinding model tailored to the lateral movement context, in which we try to find out the attacker’s location to deploy deception assets accurately for interception. The attack–defense process is modeled as a multi-agent stochastic game, by comparing it with random action policy and Minimax-Q algorithm, we choose Nash Q-learning to solve the deception asset’s deployment strategy to achieve the optimal solution effect. Extensive simulation tests reveal that our proposed model exhibits good convergence properties. Moreover, the average defense success rate surpasses 70%, attesting to the model’s efficacy.

Exploring and analyzing safety influencing factors can guide targeted traffic safety management. Traditional traffic safety models are aimed at specific data problems and making adjustments to the model structure, which lack focus on predictive ability and have limited information on the analysis of influencing factors. In recent years, machine-learning methods have opened new avenues in modeling that have higher prediction accuracy, can identify complex nonlinear relationships, and can overcome over- and under-dispersion and correlation. Machine-learning methods, however, pose the problem of limited interpretability. The interpretable machine-learning framework SHAP can be an effective solution, which can not only reflect the influence of features in each sample but also generate global interpretation. This study established gradient boosting models including the CatBoost and XGBoost models as traffic safety models, which were compared with a traditional NB regression model and a zero-inflated negative binomial regression model. SHAP was used to analyze several safety influencing factors, including geometric design features, traffic operation characteristics, time of day, and land use. Results confirmed that the CatBoost model has better prediction ability and is a more suitable traffic safety model than the traditional negative binomial regression model. Among the key findings were that ramp type is the most important factor in freeway crash frequency; curve presence has a great positive impact, while truck proportion has a great negative impact; and traffic volume is highly correlated with truck proportion. These findings can provide theoretical support for safety operation management and targeted improvement measures for freeways.

Due to the fine-grained splitting of microservices and frequent communication between microservices, the exposed attack surface of microservices has exploded, facilitating the lateral movement of attackers between microservices. To solve this problem, a multi-dimensional moving target defense method based on an adaptive simulated annealing genetic algorithm (MD2RS) is proposed. Firstly, according to the characteristics of microservices in the cloud, a microservice attack graph is proposed to quantify the attack scenario of microservices in the cloud so as to conveniently and intuitively observe the vulnerability of microservices in the cloud and the dependency relationship between microservices. Secondly, the security gain and resource cost are quantified for the key nodes selected by measuring the degree of dependence of each node according to the degree centrality. Finally, the Adaptive Simulated Annealing Genetic Algorithm (ASAGA) is used to solve the optimal security configuration information of the moving target defense, that is, the combination of the number of copies of the multi-copy deployment and the rotation cycle of the dynamic rotation of microservices, in order to quickly evaluate the security risks of microservices and optimize the security policy. Experiments show that the defense return rate of MD2RS is 85.95% higher than that of the mainstream methods, and the experimental results are conducive to applying this method to the dynamic defense of microservices in the cloud.

Truck crashes are generally more serious than passenger vehicle crashes, and they cause more deaths per crash worldwide per the U.S. Department of Transportation's Fatality Analysis Reporting System. Risk assessment and factor analysis are the keys to preventing truck crashes, but research on commercial trucks has been limited. Currently, freight and insurance companies have collected extensive operating data, now making it possible to obtain deep insights into truck crashes. Vehicle trajectory data and in-vehicle monitoring data were collected for 596 large commercial trucks traveling in Shanghai, China, during 2019. A total of 22 variables were extracted, falling into three aspects: driving behavior, travel characteristics, and warning characteristics. The random forest algorithm was used to select the most important variables for further analysis. Four machine learning models and a mixed effects logistic regression model were developed to link the high-importance variables with crash risk. Results showed that the machine learning models had good predictive performance; the bagging tree model performed best overall, having achieved good performance in the majority of the metrics, with an accuracy of 96.1% and area under the characteristic curve of 0.866. The specific variables significantly associated with crash risk were: average freeway speed, average percentage of time spent speeding, driving hours, percentage of nighttime trips, percentage of freeway trips, and frequency of smoking warnings per 100 km. This study's findings can be used to support proactive safety management for freight companies and policy formulation for insurance companies.

Adversarial distillation (AD) has emerged as a potential solution to tackle the challenging optimization problem of loss with hard labels in adversarial training. However, fixed sample-agnostic and student-egocentric attack strategies are unsuitable for distillation. Additionally, the reliability of guidance from static teachers diminishes as target models become more robust. This paper proposes an AD method called Learnable Distillation Attack Strategies and Evolvable Teachers Adversarial Distillation (LDAS&ET-AD). Firstly, a learnable distillation attack strategies generating mechanism is developed to automatically generate sample-dependent attack strategies tailored for distillation. A strategy model is introduced to produce attack strategies that enable adversarial examples (AEs) to be created in areas where the target model significantly diverges from the teachers by competing with the target model in minimizing or maximizing the AD loss. Secondly, a teacher evolution strategy is introduced to enhance the reliability and effectiveness of knowledge in improving the generalization performance of the target model. By calculating the experimentally updated target model's validation performance on both clean samples and AEs, the impact of distillation from each training sample and AE on the target model's generalization and robustness abilities is assessed to serve as feedback to fine-tune standard and robust teachers accordingly. Experiments evaluate the performance of LDAS&ET-AD against different adversarial attacks on the CIFAR-10 and CIFAR-100 datasets. The experimental results demonstrate that the proposed method achieves a robust precision of 45.39% and 42.63% against AutoAttack (AA) on the CIFAR-10 dataset for ResNet-18 and MobileNet-V2, respectively, marking an improvement of 2.31% and 3.49% over the baseline method. In comparison to state-of-the-art adversarial defense techniques, our method surpasses Introspective Adversarial Distillation, the top-performing method in terms of robustness under AA attack for the CIFAR-10 dataset, with enhancements of 1.40% and 1.43% for ResNet-18 and MobileNet-V2, respectively. These findings demonstrate the effectiveness of our proposed method in enhancing the robustness of deep learning networks (DNNs) against prevalent adversarial attacks when compared to other competing methods. In conclusion, LDAS&ET-AD provides reliable and informative soft labels to one of the most promising defense methods, AT, alleviating the limitations of untrusted teachers and unsuitable AEs in existing AD techniques. We hope this paper promotes the development of DNNs in real-world trust-sensitive fields and helps ensure a more secure and dependable future for artificial intelligence systems.

Fatigue is an important cause of traffic crashes, and effective fatigue detection models can reduce these crashes. Research has found large differences in fatigued driving performance from driver to driver, as well as a significant cumulative effect of fatigue on a given driver over time. Both sources of variation can decrease the accuracy of detection systems, but previous studies have not done enough to evaluate these differences. The purpose of this study is therefore to develop a fatigue detection model that considers individual differences and the time cumulative effect of fatigue. Data on the lateral position of the car in its lane, steering wheel movement, speed, and eye movement were collected from 22 drivers using a driving simulator with an eye-tracking system. Drivers’ subjective fatigue scores were collected using the Karolinska Sleepiness Scale. State space models (SSMs) were built to detect fatigue in each driver, considering his or her individual features. As a time series model, the SSM can also address the time cumulative effect of fatigue, and it does not require a large dataset to achieve high levels of accuracy. The differences in SSM results confirm that diversity does exist among drivers’ fatigued driving performance, so the ability of the SSM to take into account driver-specific information from each individual driver suggests that it is more suitable for fatigue detection than models that use aggregated driver data. Results show that the fatigue detection accuracy of the SSM (77.73%) is higher than that of artificial neural network models (61.37%). The advantages of accuracy, high interpretability, and flexibility make the SSM a comprehensive and valuable individualized fatigue detection model for commercial use.

Freeway crash prediction models are the basic of traffic safety research, yet crash occurrence and the influencing factors change over time. In order to make sure the implemented safety models fit the current traffic environment, this study conducts a comparative analysis of 2017 and 2020 datasets collected from freeways in Suzhou, China. Considering the spatial correlation among analysis units and the hierarchical data structure, a Bayesian conditional autoregressive negative binomial (CAR-NB) model and a Bayesian hierarchical CAR-NB (HCAR-NB) model were used to explore the safety influencing factors, and a traditional NB model was developed for further comparison. To update the HCAR-NB model from 2017 to 2020, Bayesian inference with informative priors was used to improve its goodness of fit and efficiency. Preliminary results showed that 1) the HCAR-NB model outperformed the NB model and CAR-NB model in prediction accuracy, and 2) the number of crashes was significantly correlated with average speed, speed variance, road segment length, number of lanes, and presence of ramps. The potential for safety improvement (PSI) method was applied to the modeling results to identify hotspots for the two years. The results confirmed that the hotspots spatiotemporally shifted among the freeways. The proposed crash prediction model and updating method are expected to assist implementation of informed countermeasures for freeway safety improvement.

The traffic crash rate of non-motorized vehicles (NMV) at signalized intersections is high. A thorough understanding of NMV pre-crash scenarios can lead to targeted improvement of safety, but NMV pre-crash scenarios are complex and diverse, and research is limited. Based on 1,894 NMV crashes at signalized intersections in Shanghai, China, this study identified NMV pre-crash scenarios. The characteristics and contributing factors of the NMV pre-crash scenarios were identified using association rule data mining. The results showed that the five most common scenarios were an angle collision between the motorized vehicle (MV) going straight and the NMV (representing 34.41%), followed by angle collision between the MV turning right and the NMV (13.3%), sideswipe collision between the MV going straight and the NMV traveling in the opposite direction (12.92%), sideswipe collision between the MV going straight and the NMV traveling in the same direction (11.47%), and a sideswipe collision between the MV turning right and the NMV turning in the same direction (10.88%). The most frequent contributing factors were the four-legged signalized intersection, access near the intersection, no median, bike lane, rider age between 46 and 65, rider violating traffic light, rider not giving way, clear weather, and daytime lighting. The association rule results showed, however, that the five common pre-crash scenarios varied in characteristics and contributing factors. These findings provide a unique understanding of NMV crashes in different scenarios, which can lead to well-targeted enforcement, education, and engineering-based countermeasures to solve or mitigate NMV safety.

One challenge facing image quality assessment (IQA) is that current models designed or trained on the basis of exiting databases are intrinsically suboptimal and cannot deal with the real-world complexity and diversity of natural scenes. IQA models and databases are heavily skewed toward the visibility of distortions. It is critical to understand the wider determinants of perceived quality and use the new understanding to improve the predictive power of IQA models. Human behavioral categorization performance is powerful and essential for visual tasks. However, little is known about the impact of natural scene categories (SCs) on perceived image quality. We hypothesize that different classes of natural scenes influence image quality perception—how image quality is perceived is not only affected by the lower level image statistics and image structures shared between different categories but also by the semantic distinctions between these categories. In this article, we first design and conduct a fully controlled psychovisual experiment to verify our hypothesis. Then, we propose a computational framework that integrates the natural SC-specific component into image quality prediction. Research demonstrates the importance and plausibility of considering natural SCs in future IQA databases and models.

In February 2020, the Stockholm Declaration was announced, urging states toward the United Nation’s target of a 50% reduction in traffic deaths and injuries by 2030, with the potential to achieve Vision Zero by 2050. The aim of this research is threefold namely i) to assess if selected developed countries are likely to achieve the 2030 target, ii) to use the Gompertz model to predict future road trauma trends based on historical data and iii) to understand how number of road traffic fatality vary between the developed countries. After identifying potential reasons behind the patterns, time series models were applied to identify the effects of exposure variables on traffic fatalities. To assess the likelihood of meeting the U.N. target, autoregressive integrated moving average (ARIMA) models were used for obtaining trustworthy forecasts of road traffic fatalities using data from the last five decades from seven high-income countries. The total number of fatalities, vehicle-km travelled, vehicle ownership, GDP, GDP per capita, urbanization, population density and country population were used to develop the ARIMA models. The predictive performance of the models was validated for each country, and all were found to be within the 95% confidence interval. Estimated forecasts in all seven countries appear to be realistic, with Greece and the U.K., the only countries falling short of achieving the U.N.’s 2030 target. With these results, both developed and developing countries can review and reconsider the effects of safety interventions and other socioeconomic influences on achieving reductions in road fatalities. Interventions can be added to the existing model to ascertain their effect on the predicted number of fatalities.

How best to measure spatial saliency shift induced by image distortions is an open research question. Our previous study has shown that image distortions cause saliency to deviate from their original places in natural images, and the degree of such distortion-induced saliency variation (DSV) depends on image content and the properties of distortion. Being able to measure DSV benefits the development of saliency-based image quality algorithms. In this article, we first investigate the plausibility of using existing mathematical algorithms for measuring DSV and their potential limitations. We then develop a new algorithm for quantifying DSV based on a deep neural network. In the algorithm, namely, saliency similarity transformation DSV (ST-DSV), we design a coarse-grained to fine-grained saliency similarity transformation approach to achieve DSV measurement. The experimental results show that the proposed ST-DSV algorithm significantly outperforms existing methods in predicting the ground truth DSV.

Safety performance functions (SPFs) are indispensable analytical tools that usually play a crucial role in estimating crash frequencies, identifying hotspots, analyzing crash contributing factors, and assessing the effectiveness of safety countermeasures. Due to the limited availability of safety data, municipalities tended to adopt SPFs from Highway Safety Manual or other neighboring jurisdictions. Recently, boosting algorithms have been frequently exploited for data analysis and statistical regression modeling problems. This research, therefore, aims to examine the efficiency of boosting calibration techniques to transfer the SPF using the limited region data in an international context. To this end, AdaBoost.R2, an adaptive boosting algorithm, Two-stage TrAdaBoost.R2, an instance-based transfer learning algorithm, and Gradient Boosting algorithm were employed to investigate their efficiencies in acquiring knowledge from the available source domain data to predict crashes in the target domain. As a comparison, the calibration factor method was adopted to transfer the traditional negative binomial (NB) regression model. Two training dataset groups were developed to train the four calibration techniques. The first group was used to examine the adaptability of the employed calibration techniques to the limited target region data. While the second group was utilized to further investigate the influence of larger vital information on the performance of transferred models. This study was conducted between two U.S. states, Florida and New York, and two Chinese cities, Shanghai and Suzhou. According to the goodness-of-fit results, boosting calibration techniques showed better prediction accuracy than the calibrated NB-based model using the limited target region data. In addition, the amount and distribution of the training dataset were considered the two significant factors that influence the proficiency of the boosting calibration techniques.

Drowsy driving is one of the main causes of traffic crashes, a serious threat to road traffic safety. The effective early detection of a drowsiness state can help provide a timely warning for drivers, but previous studies have seldom considered the cumulative effect of drowsiness over time. The purpose of this study is therefore to establish a model to detect a driver's drowsiness level by considering individual differences combined with the time cumulative effect (TCE) of drowsiness. Driving behavior and eye movement data from 27 drivers were collected by a driving simulator with an eye-tracking system, and the Karolinska Sleepiness Scale (KSS) was used to record drivers' perceptions of their states of drowsiness. Since the degree of driver drowsiness was shown to increase with time, a mixed-effect ordered logit (MOL) model was established, and a non-decreasing function of time was applied to consider time accumulation. Results showed that with increasing drowsiness, the standard deviation of lateral position and percentage of driver eyelid closure (PERCLOS) increased significantly. Consideration of these variables can thus improve the accuracy of drowsy driving detection. The developed MOL-TCE model was compared with a non-TCE MOL and a TCE mixed generalized ordered response (MGOR) model. The drowsiness detection accuracy of the MOL-TCE model was 62.84%, higher than the 61.04% accuracy of the MGOR-TCE and appreciably higher than the 52.47% of the non-TCE MOL model. (C) 2020 Elsevier Ltd. All rights reserved.