检索结果-内蒙古大学图书馆

FedCSpc: A Cross-Silo Federated Learning System with Error-Bounded Lossy Parameter Compression

IEEE Transactions on Parallel and Distributed Systems 2025年第7期36卷 1372-1386页

作者： Zhang, Zhaorui Di, Sheng Zhao, Kai Jin, Sian Tao, Dingwen Ji, Zhuoran Liu, Benben Alharthi, Khalid Ayed Cao, Jiannong Cappello, Franck The Hong Kong Polytechnic University Department of Computing Hong Kong Florida State University Department of Computer Science United States Temple University Department of Computer and Information Sciences United States Indiana University Bloomington Luddy School of Informatics Computing and Engineering United States Shandong University School of Cyber Science and Technology China The University of Hong Kong LSCM R&D Center Hong Kong University of Bisha Department of Computer Science College Of Computing And Information Technology Saudi Arabia Argonne National Laboratory United States

Cross-Silo federated learning is widely used for scaling deep neural network (DNN) training over data silos from different locations worldwide while guaranteeing data privacy. Communication has been identified as the main bottleneck when training large-scale models due to large-volume model parameters and gradient transmission across public networks with limited bandwidth. Most previous works focus on gradient compression, while limited work tries to compress parameters that can not be ignored and extremely affect communication performance during the training. To bridge this gap, we propose FedCSpc: an efficient cross-silo federated learning system with an XAI-driven adaptive parameter compression strategy for large-scale model training. Our work substantially differs from existing gradient compression techniques due to the distinct data features of gradient and parameter. The key contributions of this paper are fourfold. (1) Our designed FedCSpc proposes to compress the parameter during the training using the state-of-the-art error-bounded lossy compressor – SZ3. (2) We develop an adaptive compression error bound adjustment algorithm to guarantee the model accuracy effectively. (3) We exploit an efficient approach to utilize the idle CPU resources of clients to compress the parameters. (4) We perform a comprehensive evaluation with a wide range of models and benchmarks on a GPU cluster with 65 GPUs. Results show that FedCSpc can achieve the same model accuracy as FedAvg while reducing the data volume of parameters and gradients in communication by up to 7.39× and 288×, respectively. With 32 clients on a 4Gb size model, FedCSpc significantly outperforms FedAvg in wall-clock time in the emulated WAN environment (at the bandwidth of 1 Gbps or lower without loss of generality). © 1990-2012 IEEE.

关键词： Bandwidth compression

来源：评论

学校读者我要写书评

暂无评论

BreastLightSAM: A Lightweight Pipeline for Fast and Accurate Breast Cancer Diagnosis and Tumor Segmentation

BreastLightSAM: A Lightweight Pipeline for Fast and Accurate...

引用

Medical Imaging 2025: Ultrasonic Imaging and Tomography

作者： Hu, Mingzhe Yang, Xiaofeng Department of Radiation Oncology Winship Cancer Institute Emory University AtlantaGA30322 United States Department of Computer Science and Informatics Emory University AtlantaGA30322 United States

ISBN: (纸本)9781510686021

BreastLightSAM presents an innovative, lightweight pipeline tailored to tackle the challenges associated with rapid and precise diagnosis of breast cancer in 2D ultrasound images. The model integrates a modified Segmentation Anything Model (SAM) with an optimized RepViT-based image encoder for efficient segmentation and incorporates a compact classification module based on GhostNet to accurately classify breast tumors. This design allows for high segmentation accuracy and classification precision while significantly reducing latency, making it suitable for deployment on computationally limited devices like mobile phones. The proposed pipeline features a point-based prompt design that emulates human interaction, enhancing the model's generalization and robustness. Results show that BreastLightSAM achieves a mean Dice Similarity Coefficient (DSC) of 0.812 and a mean pixel accuracy of 0.943, with a latency of just 28.6 milliseconds, while the classification module delivers high precision and recall across tumor types. However, the study's focus on 2D ultrasound images limits its exploration of the impact of 3D imaging on performance. Additionally, the influence of human expertise and pre-processing techniques such as contrast enhancement and image sharpening on segmentation results requires further investigation. These aspects present valuable avenues for future research to enhance the clinical applicability of BreastLightSAM in diverse healthcare settings. © 2025 SPIE.

关键词： Image segmentation

来源：评论

学校读者我要写书评

暂无评论

Towards Robust Federated Learning: Investigating Poisoning Attacks Under Clients Data Heterogeneity 19

Towards Robust Federated Learning: Investigating Poisoning A...

引用

19th International Conference on Ubiquitous Information Management and Communication, IMCOM 2025

作者： Soubih, Abdenour Lahmer, Seyyid Ahmed Abuhamad, Mohammed Abuhmed, Tamer College of Computing and Informatics Sungkyunkwan University Suwon Korea Republic of Department of Information Engineering University of Padova Padua Italy Department of Computer Science Loyola University Chicago Chicago United States

ISBN: (纸本)9798331507817

Federated Learning (FL) offers a privacy-preserving solution by enabling multiple clients to train a shared model collaboratively without centralizing data. However, the decentralized nature of FL presents challenges, particularly regarding security and performance under adversarial conditions. This paper investigates the effects of poisoning attacks under data heterogeneity. Our experiments evaluate the impact of varying malicious client fractions and poison concentration levels on the accuracy of the model. We explore the effects of poisoning attacks on FedAvg and FedNova models using medical imaging tasks. Our findings reveal that increasing data heterogeneity exacerbates the effects of poisoning, with FedNova demonstrating greater resilience compared to FedAvg. We found that the number of malicious clients plays a more significant role in degrading performance than the ratio of poisoning samples shared by each malicious client, suggesting that even modest levels of poisoning can be tolerated by most algorithms. The study highlights the importance of developing robust defense mechanisms to maintain model performance under adversarial conditions. © 2025 IEEE.

关键词： Federated learning

来源：评论

学校读者我要写书评

暂无评论

AI-Driven Innovation Using Multimodal and Personalized Adaptive Education for Students With Special Needs

引用

IEEE Access 2025年 13卷 69790-69820页

作者： Farhah, Nesren Adnan, Muhammad Abdullah Alqarni, Ahmed Irfan Uddin, M. Aldhyani, Theyazn H. H. King Salman Center for Disability Research Riyadh11614 Saudi Arabia Saudi Electronic University College of Health Science Department of Health Informatics Riyadh11673 Saudi Arabia Institute of Computing Kohat University of Science and Technology Kohat26000 Pakistan Al-Baha University Department of Computer Sciences and Information Technology Al-Aqiq 65779 Saudi Arabia Al Hofuf31982 Saudi Arabia

This study provides an in-depth exploration of the use of multimodal techniques in developing adaptive learning systems designed for students with special needs using various neural network models: a Baseline Neural Network, Convolutional Neural Network, Attention Model, LSTM, GRU, and Transformer models. Adaptive learning systems play a major role in the customization of the educational process in such a way as to ensure the specific cognitive, behavioural and emotional needs of students with special educational needs are catered for. The models were trained using a sizeable amount of multimodal data that recorded factors such as academic achievement, behavioral interactions, and environmental factors. From our observations, it can be noted that transformer models and attention mechanisms allowed for higher performance in predicting and fulfilling the needs of students, and were more effective than simpler neural architectures. These findings suggest that more advanced and intelligent neural networks can improve how students with different learning abilities consume content and, consequently, provide better educational answers and solutions to the market. To identify the important features affecting student performance, various Explainable AI (XAI) techniques were utilized, including SHAP (Shapley Additive Explanations) values, logistic regression-based feature importance analysis, LIME (Local Interpretable Model-Agnostic Explanations), partial dependency plots, and whose global surrogate model was a Decision Tree Regressor. These XAI instruments allowed the model’s decision to be explained and visualized so that teachers could understand the main beneficiaries of the improved educational results. Such XAI methods enable transparency in how models make decisions, thus allowing educators to understand the reasons for certain recommended learning paths. © 2013 IEEE.

关键词： Convolutional neural networks

来源：评论

学校读者我要写书评

暂无评论

Enhancing Internet Traffic Forecasting in MEC Environments With 5GT-Trans: Leveraging Synthetic Data and Transformer-Based Models

引用

IEEE Access 2025年 13卷 83607-83618页

作者： Shukla, Ravi Shankar Alatawi, Abdulrahman M. Saudi Electronic University College of Computing and Informatics Department of Computer Science Riyadh 11673 Saudi Arabia Saudi Electronic University Department of Computer Science Riyadh 11673 Saudi Arabia

Recently, transformer-based learning models have shown promising results in network traffic prediction across various contexts. However, their potential in Mobile Edge Computing (MEC) environments remains largely unexplored, despite the unique requirements of MEC such as low latency and high-bandwidth communication between end-users and edge servers. This study evaluates the performance of four transformer-based learning models for Internet traffic prediction in MEC environments, comparing them against state-of-the-art models using a real-world dataset from publicly available mobile operator networks. Performance is assessed using mean squared error (MSE) and mean absolute error (MAE) as evaluation metrics. Our experiments demonstrate that the proposed GAN-based transformer model, referred to as 5GT-GAN-Trans, achieves an MSE of 0.428 and an MAE of 0.441, considerably outperforming baseline models such as FED-Former-F (MSE: 0.791, MAE: 0.606), FED-Former-W (MSE: 0.622, MAE: 0.508), Informer (MSE: 0.814, MAE: 0.649), and a standard Transformer model (MSE: 0.803, MAE: 0.640). These numerical improvements underscore the efficacy of transformer-based models for precise Internet traffic forecasting in MEC environments, offering valuable insights for improved network management and enhanced user experience. © 2013 IEEE.

关键词： 5G generative adversarial network Internet traffic prediction mobile edge computing

来源：评论

学校读者我要写书评

暂无评论

Optimization of 6G resource allocation using CyberTwin function-based service enhancement scheme

引用

Eurasip Journal on Wireless Communications and Networking 2025年第1期2025卷 1-27页

作者： Aldrees, Asma Min, Hong Daradkeh, Yousef Ibrahim Dutta, Ashit Kumar Anjum, Mohd Department of Informatics and Computer Systems College of Computer Science King Khalid University Abha61421 Saudi Arabia School of Computing Gachon University Seongnam13120 Korea Republic of Department of Computer Engineering and Information College of Engineering in Wadi Alddawasir Prince Sattam Bin Abdulaziz University Al-Kharj16273 Saudi Arabia Department of Computer Science and Information Systems College of Applied Sciences Almaarefa University Ad Diriyah Riyadh13713 Saudi Arabia Department of Computer Engineering Aligarh Muslim University Aligarh202002 India

The most innovative Service-Optimized Logging for Resource Allocation (SOL-RA) is in situations that support 6G connection. The CyberTwin architecture optimizes the quality of service provided to end users in 6G communications by using the dependability of terahertz connections and interactions of a machine-type nature. Elastically sharing elastic resources among users is made possible by the proposed resource allocation strategy, which makes use of proprietary logger functions. This ensures that efficient allocation is achieved without overlapping or lengthy wait periods. Within the framework of SOL-RA, a categorization mechanism is implemented for dense requests, which differentiates them as either stationary or priority services. Resource allocations are dynamically done based on this categorization, which enhances the organization's responsiveness to different needs. Through a tree classifier learning mechanism, CyberTwin logs play an essential part in the processing of requests and the preparation of resources. This helps to ensure that resources are distributed without inefficiency. Requests that are generating a backlog are identified and processed in parallel by the system, which significantly reduces the amount of time that is wasted waiting. In order to guarantee a specialized distribution for the categorized outputs, resource allocations are directed by CyberTwin data logs. An examination of the performance of the SOL-RA scheme takes into account important metrics such service latency, service backlog, resource usage, and request-to-response ratio. This research offers insights into the efficacy of the scheme in maximizing service quality and resource consumption in 6G settings. © The Author(s) 2025.

关键词： Resource allocation

来源：评论

学校读者我要写书评

暂无评论

Deep Reinforcement Learning-Based Framework for Enhancing Cybersecurity

International Journal of Interactive Mobile Technologies

引用

International Journal of Interactive Mobile Technologies 2025年第3期19卷 170-190页

作者： Al-Nawashi, Malek M. Al-Hazaimeh, Obaida M. Tahat, Nedal M. Gharaibeh, Nasr Abu-Ain, Waleed A. Abu-Ain, Tarik Al-Huson University College Al-Balqa Applied University Irbid Jordan Faculty of Information Technology Department of Information Security and Cybersecurity Philadelphia University Amman Jordan Faculty of Science The Hashemite University Zarqa Jordan College of Science and Computer Engineering Taibah University Yanbu Saudi Arabia College of Computing and Informatics Saudi Electronic University Riyadh Saudi Arabia

The detection of cyberattacks has been increasingly emphasized in recent years, focusing on both infrastructure and people. Conventional security measures such as intrusion detection, firewalls, and encryption are insufficient in protecting cyber systems against growing and changing threats. In order to address this problem, scholars have explored reinforcement learning (i.e., RL) as a potential solution for intricate cybersecurity decision-making difficulties. Nevertheless, the use of RL faces several obstacles, including dynamic attack scenarios, insufficient training data, and the challenge of replicating real-world complexities. This study presents a novel framework that uses deep reinforcement learning (i.e., DRL) to simulate harmful cyberattacks and improve cybersecurity. This study presents an agent-based framework that is capable of ongoing learning and adaptation in a dynamic network security environment. The agent determines the optimal course of action by considering the current state of the network and the rewards it receives for its decisions. The CIC-IDS-2018 database, constructed using Python 3.7 programming, was used. The conducted studies yielded outstanding results, with a detection accuracy of 98.82% achieved for the CIC-IDS-2018 database in cyber-attack classification. © 2025 by the authors of this article. Published under CC-BY.

关键词： Cybersecurity

来源：评论

学校读者我要写书评

暂无评论

AtheroRisk: An Integrated computer Software System for Stroke Risk Stratification Utilizing Carotid Plaque Analysis in Ultrasound Videos

引用

SN computer science 2025年第2期6.0卷 1-16页

作者： Gemenaris, Michalis Liapi, Georgia D. Markides, Christos Constantinou, Kyriacos Loizou, Christos P. Neophytou, Michalis Kardoulas, Dimitrios Pattichis, Marios S. Griffin, Maura Nicolaides, Andrew Pattichis, Constantinos S. Kyriacou, Efthyvoulos Department of Electrical Engineering Computer Engineering and Informatics Cyprus University of Technology Limassol Cyprus Department of Computer Science and Engineering Frederick University Nicosia Cyprus Department of Computer Science University of Cyprus Nicosia Cyprus American Medical Center Nicosia Cyprus Vascular Surgery Clinic Attikon General University Hospital Athens Greece Department of Electrical and Comp. Engineering Univ. of New Mexico Albuquerque NM United States Vascular Screening and Diagnostic Center Nicosia Cyprus

In this study we propose and evaluate AtheroRisk, a standalone integrated computer software system for the analysis of carotid B-mode ultrasound (U/S) images and videos. Our goal was to provide a tool to help physicians stratify stroke risk. The presented system is based on outcomes from different research studies, as well as European guidelines for the evaluation and treatment of carotid artery disease. The goal is to enable reproducible analysis of atherosclerotic plaques in real time. AtheroRisk brings together analysis of U/S image and/or video of the carotid arteries facilitating anonymization, standardization, noise filtering and segmentation of plaques;followed by image and motion feature extraction to derive plaque-composition and stability. All the above steps are combined to determine the annual stroke risk rate and the 5-year stroke-free survival rate. The system is powered by an SQLite local database, in which the end user can save and manage data extracted from processing and analysis in an anonymized way. The first version of AtheroRisk was developed following an incremental software development process. The developed system was verified using 54 U/S videos (27 Asymptomatic, AS;27 Symptomatic, SY cases) while the clinician’s satisfaction and comments were collected through a questionnaire-based validation process. The results show that the integrated system proposed in this study can be successfully used for the automated image and video analysis of the CCA plaque in ultrasound videos. © The Author(s) 2025.

关键词： Carotid plaque analysis Motion analysis Plaque features Stroke risk stratification Ultrasound video

来源：评论

学校读者我要写书评

暂无评论

TCIP: Transformed Colour Intensity Profiles analysis for fault detection in power line insulators

引用

Multimedia Tools and Applications 2025年第8期84卷 4503-4534页

作者： Tomaszewski, Michal Gasz, Rafal Kasana, Singara Singh Osuchowski, Jakub Singh, Simranjit Zator, Slawomir Department of Computer Science Faculty of Electrical Engineering Automatic Control and Informatics Opole University of Technology Prószkowska 76 Opole45-758 Poland Department of Computer Science and Information Technology Central University of Haryana Haryana Mahendergarh123031 India Department of Information Technology Dr B. R. Ambedkar National Institute of Technology Punjab Jalandhar India Department of Safety Engineering and Technical Systems Faculty of Production Engineering and Logistics Opole University of Technology Prószkowska 76 Opole45-758 Poland

In response to the escalating demand for electricity, the aging process and inherent failures in power lines have become unavoidable challenges in their operational integrity. This research addresses the imperative need for a robust method to detect damage in power insulator disc images, introducing a novel feature extraction technique grounded in color intensity profile transformation and ensemble learning. For the purposes of the research, a specific dataset was developed based on publicly available images of power insulators. The collection consists of 9232 images of insulator discs which included 2184 images of damaged discs. The discernible damages include breakages and burnouts. Two distinct analysis variants were employed. In the first variant (V1), two classification groups were selected: disc with damages and disc without damages. In the second variant (V2), three classification groups were selected: discs without damage, broken discs, and burnout discs. The described research was divided into two successive stages. In the first of them (S1), the transformed color intensity profiles for the prepared data set were built and various machine learning models for their classification were trained and evaluated. In the next stage (S2), previously developed, highest efficient machine learning models, were used to make the detection of various damaged types of overhead line insulator discs. For the first stage of research high-performance results and a high value of the F1 score parameters were obtained (F1 score = 0.999 for the first variant (S1V1) and F1 score = 0.992 for the second variant (S1V2). In the second stage of research and the first variant (S2V1), the highest result of the F1 score was 0.932 and for the second variant (S2V2) F1 score was 0.930. © The Author(s), under exclusive licence to Springer science+Business Media, LLC, part of Springer Nature 2024.

关键词： Feature extraction

来源：评论

学校读者我要写书评

暂无评论

Unlocking Domain Specificity: Fine-Tuning Llama 2 for Enhanced Performance on Custom Datasets

Unlocking Domain Specificity: Fine-Tuning Llama 2 for Enhanc...

引用

International IoT, Electronics and Mechatronics Conference, IEMTRONICS 2024

作者： Bhattacharya, Swarnadwip Bhattacharjee, Anindita Das Roy, Pranab Singha Samanta, Tapas Department of Computer Science and Engineering Institute of Engineering and Management Kolkata700091 India Department of CSE IEM Centre of Excellence for InnovAI Institute of Engineering and Management Kolkata700091 India Department of Computer and Informatics Variable Energy Cyclotron Centre Kolkata700064 India

ISBN: (纸本)9789819747795

This paper aims to improve Llama 2’s performance by using personalized and modified datasets. Despite the impressive capabilities of large language models (LLMs) such as Llama 2, their effectiveness may be limited in specialized domains. The proposed method entails fine-tuning Llama 2 on custom datasets to optimize performance efficiently. The study focuses on the impact of quantization-aware low-rank adapter layers (QLoRA) on a single GPU’s resource-efficient fine-tuning performance. Furthermore, the study looks into the design of instruction datasets to guide the model toward desired behaviors. When Llama 2 is fine-tuned with QLoRA, performance improves significantly across tasks such as text summarization, question answering, and natural language generation in a variety of domains. The paper concludes by highlighting the broader implications of the findings. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

关键词： Custom datasets Fine-tuning Instruction datasets Large language models (LLMs) Llama 2 Personalized datasets QLoRA Quantization-aware low-rank adapter layers (QLoRA) Resource-efficient fine-tuning Specialized domains

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：