Recognizing vehicles under difficult environmental conditions is an enduring challenge to the effective implementation of intelligent transportation systems and autonomous driving technologies. In everyday life, many factors, including poor weather (rain), night-time illumination & darkness, and partial occlusion will greatly reduce the reliability of typical deep learning-based vehicle identifiers. This paper proposes a robust framework for vehicle recognition using 3 complementary strategies: data augmentation; domain adaptation using a Domain- Adversarial Neural Network (DANN); and applying an image enhancement module based on Zero-Reference Deep Curve Estimation (Zero-DCE). The UA-DETRAC dataset was used to develop a robust evaluation protocol based on 5 different scene types: sunny, rainy, night-time, low-light and occluded. At the validation stage, both ResNet50 and EfficientNet-B0 were used as baseline models. Experiments were carried out to compare each of the 3 strategies both separately and in combination. The results showed that using the image enhancement module produced the greatest improvement (8.3% & 7.9%) in coordination with the low-light & night-time scene types respectively. Domain adaptation has improved recognition performance (i.e., the accuracy of recognition) for almost every scene category (particularly those with occlusions), where the use of domain adaptation has increased recognition performance by 6.1%. In addition, combining all three approaches resulted in a very large overall performance increase from an average of 73.4% (baseline) to 87.6%, an improvement of 14.2%. The paper also looks at the time that each method takes to make a prediction and how they can be used together to improve vehicle recognition in applications.

As a classic strategy confrontation board game, chess features both complex spatial layout characteristics and long-term sequential dependency, making it an important research scenario in the fields of artificial intelligence and machine learning. This study is based on 3,196 endgame match data and 35 board state features, constructing a multi-model comparison framework including decision trees, logistic regression, Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNN), to verify the performance differences of different algorithms in the task of predicting the outcome of chess moves. Through feature encoding, dataset division, and multi-model training and evaluation, the strengths and weaknesses of each model in terms of accuracy, generalization ability, and inference efficiency are analyzed. Preliminary experiments show that logistic regression as a baseline model can effectively learn simple chess game patterns. Subsequently, the introduction of LSTM and CNN revealed that the dataset is relatively weak in sequence modeling and spatial feature extraction, resulting in low accuracy. Further improvement in prediction accuracy is needed. This research provides a reference basis for the model selection of the intelligent analysis system for board games, and also offers practical ideas for the prediction task that integrates temporal and spatial features.

An escalation in sophisticated, encrypted botnet attacks has accompanied the rapid expansion of the Internet of Things (IoT) ecosystem. Conventional security measures are increasingly ineffective due to the rise of encrypted communication channels, while standard machine learning and deep learning models are computationally expensive for local edge deployment. This paper first proposes a hybrid, dual-branch deep learning framework designed for high-fidelity botnet detection within containerized edge environments. Experimental evaluations are then conducted using a curated dataset derived from IoT-23 and tested within a Docker-based simulation. The proposed framework in the resource-constrained setting can achieve high accuracy and a low False Positive Rate while maintaining superior computational efficiency and outperforming traditional baseline classifiers. These findings demonstrate strong capabilities for robust detection at the network edge with negligible memory overhead, offering a scalable solution for modern IoT infrastructures. Finally, the paper provides a critical assessment of the framework's limitations and future research directions of botnet detection in edge deployment scenarios.

Artificial intelligence has become a key tool in quantitative trading, while the multi-factor model remains a common framework for stock selection and asset allocation. Existing reviews, however, usually discuss multi-factor models or AI-based quantitative strategies separately. They rarely provide a systematic account of how the two approaches are integrated. This paper aims to fill this gap by reviewing recent Chinese and international studies and examining the integration of multi-factor models with artificial intelligence in quantitative trading. The review shows that traditional multi-factor models provide interpretable variables and economic logic, but they have limitations in factor construction, nonlinear modelling and market regime adaptation. Machine learning improves factor screening and return prediction; deep learning extends the use of time-series and textual information; reinforcement learning further links prediction results with portfolio adjustment and trading decisions. These developments increase the flexibility of quantitative strategies, but they also bring problems such as overfitting, weak interpretability, unstable generalization and regulatory risk. Future research should give more attention to explainable AI, dynamic factor selection, multi-source data fusion and risk-constrained trading systems. Such work can help connect economic meaning with data-driven modelling and improve the practical value of AI-enhanced multi-factor strategies.

With quantum computing now widely treated as a credible adversary to conventional public-key assumptions—most notably discrete logarithms and integer factorization—post-quantum cryptography has moved from a speculative research topic to an active engineering agenda. Within that broader landscape, lattice-based cryptography is routinely singled out as the leading candidate for next-generation digital signatures, largely because its security can be reduced to well-studied lattice problems believed resistant to quantum attacks. Aggregate signatures compress multiple individual signatures into a single compact object, translating into improved batch-verification throughput in settings such as blockchain and IoT. Against this backdrop, the present investigation offers a detailed analysis of lattice-based aggregate signature constructions. Core lattice notions are reviewed alongside security models, followed by an examination of two aggregation paradigms: unordered and ordered. Representative schemes are analyzed with respect to their construction logic and structural bottlenecks. A cross-scheme comparison is then used to motivate future research directions. By pulling together recent progress while isolating persistent obstacles, this investigation characterizes an ongoing shift from feasibility-oriented prototypes toward deployment-minded designs, offering a focused reference for standardization efforts targeting post-quantum-secure aggregate signatures.

This study examines the relationship between digital governance implementation, fiscal transparency, and economic resilience across 34 OECD countries from 2015 to 2023. Using panel data regression analysis and machine learning techniques, we find that countries with higher e-government development indices demonstrate significantly greater fiscal transparency scores (β = 0.42, p < 0.01) and improved economic resilience during global crises. The analysis reveals that a 10-percentage-point increase in digital public service adoption correlates with a 6.8% reduction in corruption perception index and a 4.2% improvement in government effectiveness ratings. Furthermore, our predictive models using random forest algorithms achieve 85.3% accuracy in forecasting fiscal transparency levels based on digital infrastructure indicators. These findings suggest that strategic investments in digital governance infrastructure serve as critical policy instruments for enhancing public sector accountability and economic stability in the post-pandemic era. The study contributes to public administration theory by providing empirical evidence for the digital-transparency-governance nexus and offers practical implications for policymakers seeking to leverage data analytics for improved public sector performance.

Demands for 3D contents across different industries are increasing such as gaming, architecture and visual arts. Recently with the development of Large Language Models (LLMs) new avenues have opened for automating and simplifying this process.This paper presents a review of current progress and applications of LLMs in 3D modeling. They can be basically categorized as two approaches: code-driven procedural 3D modeling and multimodal direct 3D synthesis. Code-driven methods like 3D-GPT and BlenderLLM, uses LLMs to translate natural language into executable scripts, which enables automated scene generation and asset editing. However these methods often lack spatial awareness. Despite significant progress, challenges such as execution errors and limited spatial reasoning remain. Finally, this paper explores future prospects, highlighting the potential of Vision-Language-Action (VLA) feedback loops, native 3D tokenization, and the shift towards synergistic human-AI co-creation workflows. These advancements may greatly improve the efficiency, accessibility, and creativity of future 3D content generation systems.

Defects in photovoltaic (PV) cells directly reduce photoelectric conversion efficiency. Automated defect detection using machine vision has become a core technology for quality control in PV manufacturing lines. This paper systematically reviews the research progress and engineering applications of machine vision-based defect detection for PV cells. First, typical defect types, formation mechanisms, and quantified impacts on PV cell performance are presented, and the principles and applicable scenarios of different imaging modalities are analyzed. Reasonable selection of imaging modalities and detection algorithms is crucial to improving detection stability and applicability in industrial scenarios. The technical principles and characteristics of traditional machine vision and deep learning methods are summarized in detail. Then, practical cases are used to compare different technical routes in terms of accuracy, real-time performance, and cost. Key solutions are proposed for challenges including lighting interference, limited defect samples, and real-time edge deployment. Finally, future trends are discussed, including vision large models, multimodal fusion, and integration with intelligent manufacturing.

In the post-Moore's Law era, microprocessor performance growth continues to slow, as the traditional approach of process scaling and frequency increases has reached a bottleneck. Physical limits, power constraints, and architectural innovation have become the three core factors influencing performance growth. However, academia disagrees on quantifying their individual contributions and lacks a unified framework for analyzing their coupled effects. This paper systematically reviews classical theories and research paradigms related to microprocessor performance growth. The study identifies three core consensuses: power constraints are the primary limiting factor throughout the performance growth cycle; architectural innovation is the key driver in the post-Moore era; and the contribution of process scaling shows clear diminishing returns. The study further compares the mechanisms of physical limits and power constraints with quantitative research on architectural innovation. By examining mainstream quantitative decomposition methods for these three factors, it highlights core shortcomings in existing research and outlines future directions. This paper provides a theoretical foundation for research on microprocessor architectural innovation and power optimization in the post-Moore's Law era.

With the rapid development of intelligent robots, their use in aid in the fields of medicine, home services, and industry collaboration has grown steadily. This has focused most of the research on ensuring safe, secure, and efficient interaction between robots, humans, and the environment. Among multiple interaction safety mechanisms, compliance control is targeting the safe and injury-free interaction of robots with people and the environment by modulating the velocity and force of robots during the interaction. As an important supportive system, multimodal perception fusion, integrates visual, tactile, and force information to predict contact states and/or assist in making compliant control decisions. The value in several applications, such as medical intervention, industrial assembly, and service robotics, is described in this paper, to explain the use case of compliance control, and to provide a foundation of intelligence embodied research on intelligent robotic interaction technology implementation of compliance control in smart robotic interaction technologies.