International Journal of Intelligent Networks最新文献_第3页

Global representation fine-tuning for federated self-supervised representation learning 联邦自监督表示学习的全局表示微调

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-09-23 DOI: 10.1016/j.ijin.2025.09.002

Hongzi Li , Guifen Zhang , Qinchun Su , Lina Ge

{"title":"Global representation fine-tuning for federated self-supervised representation learning","authors":"Hongzi Li , Guifen Zhang , Qinchun Su , Lina Ge","doi":"10.1016/j.ijin.2025.09.002","DOIUrl":"10.1016/j.ijin.2025.09.002","url":null,"abstract":"<div><div>Federated self-supervised representation learning combines federated learning with self-supervised mechanisms to learn general representations from distributed unlabeled data, effectively reducing reliance on labeled data. However, under data heterogeneity, existing methods primarily focus on aligning local and global models in the parameter space, often overlooking the issue of knowledge forgetting in the global model caused by incompatibility in local representation spaces. This limits the quality of representations and overall model performance. To address this challenge, we propose FedGRF, a global representation fine-tuning method for federated self-supervised representation learning. FedGRF maintains a generator on the server to produce pseudo-data, which is then used to drive global representation fine-tuning and mitigate the forgetting of local representation knowledge by the global model. By mining hard samples arising from the fusion of local representations and employing a controllable fine-tuning mechanism, FedGRF effectively promotes the transfer of local representation knowledge to the global model. Extensive experimental results demonstrate that FedGRF achieves competitive performance improvements over existing methods.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 224-232"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145361915","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Online and offline collaborative abnormal traffic intelligent detection system based on elastic lightweight width learning algorithm 基于弹性轻量宽度学习算法的线上线下协同异常交通智能检测系统

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-04-09 DOI: 10.1016/j.ijin.2025.04.002

Yu Wang, Hong Huang

{"title":"Online and offline collaborative abnormal traffic intelligent detection system based on elastic lightweight width learning algorithm","authors":"Yu Wang, Hong Huang","doi":"10.1016/j.ijin.2025.04.002","DOIUrl":"10.1016/j.ijin.2025.04.002","url":null,"abstract":"<div><div>As the boost of information technology, network security issues have been increasingly prominent. Therefore, it is crucial for maintaining network security to establish an efficient abnormal traffic detection system. The study first explained the width learning algorithm, which was used as the basic framework to introduce the elastic lightweight and gated neural networks for optimization. Finally, an online abnormal traffic detection model and an offline abnormal traffic detection model were proposed. The experimental results showed that the fastest iteration of the online detection model was 190, the prediction accuracy was 96 %, the prediction error floated only between −0.01 and 0.01, and the shortest computing time was 2.012 s. The minimum iteration for the offline detection model was 200, with the abnormal flow detection error of 0.11. The lowest average absolute percentage error was 0.141 and the normalized root mean square error was 0.207. The lowest root mean square error reached 0.175, and the highest R<sup>2</sup> error was 0.884. In summary, the two proposed models have achieved significant improvements in the accuracy and efficiency of abnormal traffic detection, providing a feasible solution for network security.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 27-35"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143943164","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A Comprehensive Survey on Emerging AI Technologies for 6G Communications: Research Direction, Trends, Challenges, and Opportunities 6G通信新兴人工智能技术综合调查：研究方向、趋势、挑战与机遇

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-07-07 DOI: 10.1016/j.ijin.2025.06.001

Barsha Rani Das , Syed Rakib Hasan , Saifur Rahman Sabuj , Md Akbar Hossain , Sayan Kumar Ray

{"title":"A Comprehensive Survey on Emerging AI Technologies for 6G Communications: Research Direction, Trends, Challenges, and Opportunities","authors":"Barsha Rani Das , Syed Rakib Hasan , Saifur Rahman Sabuj , Md Akbar Hossain , Sayan Kumar Ray","doi":"10.1016/j.ijin.2025.06.001","DOIUrl":"10.1016/j.ijin.2025.06.001","url":null,"abstract":"<div><div>Technology is now embedded in every aspect of human existence. The demand for technological advancement is increasing day by day for seamless connections. Integrating Artificial Intelligence (AI) into technology will make the future world beyond one's imagination. The field of wireless communication is rapidly changing, and the existing wireless networks need to be improved with the growing data rate. Fifth-generation (5G) networks are now implemented throughout the world, and they meet the criteria of modern technology with higher data rates, improved quality of service, lower latency, and so forth. However, with the changing data rates after ten years, it will become impossible to meet the demand. It is expected that the upcoming sixth-generation (6G) network with the integration of AI will be the successor of 5G, and it will resolve the main issues of 5G, like higher data rates, ultra-low latency, efficiency, and so on. AI in 6G will provide a more secure, reliable communication system than its predecessor generation. This paper presents the use of AI in 6G communication networks, technologies, techniques, trends, and future research directions. It also describes in-depth knowledge of the AI-integrated 6G communication system model. Furthermore, this paper encompasses applications of AI in several technologies, challenges, and significant developments.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 113-150"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144831555","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

MIDES: A multi-layer Intrusion Detection System using ensemble machine learning MIDES：使用集成机器学习的多层入侵检测系统

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-09-15 DOI: 10.1016/j.ijin.2025.09.001

Vincenzo Agate, Alessandra De Paola, Pierluca Ferraro, Giuseppe Lo Re

{"title":"MIDES: A multi-layer Intrusion Detection System using ensemble machine learning","authors":"Vincenzo Agate, Alessandra De Paola, Pierluca Ferraro, Giuseppe Lo Re","doi":"10.1016/j.ijin.2025.09.001","DOIUrl":"10.1016/j.ijin.2025.09.001","url":null,"abstract":"<div><div>In recent years, as the frequency and types of network attacks increase, Intrusion Detection Systems (IDSs) have become essential components of most organizations’ security infrastructure. Although the use of machine learning methods shows great promise for the design of effective IDSs, existing methods still have several limitations. Single classifiers are never able to recognize all types of attacks, regardless of the underlying algorithm. This paper proposes MIDES, a novel multi-layer IDS that integrates binary, multi-class, and meta-classifiers into a flexible architecture. MIDES employs a fast binary classifier to filter clearly benign traffic, an ensemble of specialized multi-class classifiers to analyze suspicious events, and a meta-classification layer to refine decisions. A self-adaptive agent dynamically selects the most appropriate decision strategy for each input using both static and dynamic heuristics. The system is designed to be extensible, adaptable to evolving threats, and efficient in real-time scenarios. The proposed system has been extensively evaluated on the well-known CIC-IDS2017 and CSE-CIC-IDS2018 public datasets and compared against state-of-the-art works, showing that MIDES achieves high accuracy across all 14 attack classes while significantly reducing classification time, outperforming the compared systems.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 204-223"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145099053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Designing a novel network anomaly detection framework using multi-serial stacked network with optimal feature selection procedures over DDOS attacks 设计了一种基于多串行堆叠网络的网络异常检测框架，并对DDOS攻击进行了最优特征选择

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2024-11-29 DOI: 10.1016/j.ijin.2024.11.001

K. Jeevan Pradeep, Prashanth Kumar Shukla

{"title":"Designing a novel network anomaly detection framework using multi-serial stacked network with optimal feature selection procedures over DDOS attacks","authors":"K. Jeevan Pradeep, Prashanth Kumar Shukla","doi":"10.1016/j.ijin.2024.11.001","DOIUrl":"10.1016/j.ijin.2024.11.001","url":null,"abstract":"<div><div>- Distributed denial-of-service (DDoS) attacks are the major threat that disrupts the services in the computer system and networks using traffic and targeted sources. So, real-world attack detection techniques are considered an important element in executing cybersecurity tasks. The present DDoS techniques are prone to False Positive Rates (FPR) and also it didn't acquire the complicated patterns presented in the attack traffic. Internet of Things (IoT) is a complicated network with resource-constrained devices and networks that are prone to different security threats like DDoS attacks. Later, the Software Defined Networking (SDN) with IoT models is used to enhance the access control techniques and security models. DDoS attacks are considered as an important threat in the IoT networks. Hence, it is important to construct a novel network anomaly detection model with a deep learning mechanism to resolve the limitations of the existing techniques. Initially, essential data required for the validation are gathered from the IDS ISCX 2012 dataset. The optimal features are selected from input data using the Predefined-Mud Ring Algorithm (P-MRA). The optimally selected features are provided to the Multi-Serial Stacked Networks (Multi-SSN), which is the fusion of Convolutional Autoencoder (CAE), Gated Recurrent Unit (GRU), and Bayesian Learning (BL) networks. Here, the essential features for the validation are acquired from the CAE and GRU. Then, these features are stacked and given to the BL mechanism for detecting the anomalies in the network. Further, several experimental validations are performed in the developed framework over traditional network anomaly detection mechanism.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 1-13"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143176291","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

HCLDR: Heterogeneous Graph Convolutional Networks with Contrastive Learning for Diversity Recommendation 基于对比学习的异构图卷积网络的多样性推荐

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-11-20 DOI: 10.1016/j.ijin.2025.11.007

Hanwen Liu , Jiale Xu , Shunmei Meng , Qianmu Li , Xianjin Fang

{"title":"HCLDR: Heterogeneous Graph Convolutional Networks with Contrastive Learning for Diversity Recommendation","authors":"Hanwen Liu , Jiale Xu , Shunmei Meng , Qianmu Li , Xianjin Fang","doi":"10.1016/j.ijin.2025.11.007","DOIUrl":"10.1016/j.ijin.2025.11.007","url":null,"abstract":"<div><div>Accuracy, now powering most of supporting users to achieve personalized recommendation in recommender systems, is one of the predominantly evaluation criteria. However, the diversity of the recommendation lists is limited and cannot meet the users’ more sophisticated search needs. Additionally, the performance of recommender systems is usually impaired by a highly skewed distribution of long-tail items. Indeed, manipulating the inclusion of long-tail items in the recommendation task may introduce additional noise into the learning of the user/item representations, especially for existing diversity recommendation models based on user–item heterogeneous graphs. To overcome these limitations, we propose a model called the Heterogeneous Graph Convolutional Networks with Contrastive Learning for Diversity Recommendation (HCLDR). Specifically, based on the existing user–item heterogeneous graph, a new user–item heterogeneous graph is formed by acquiring new links among items through their popularity; the embedding information of users and items is then learned based on the Heterogeneous Graph Convolutional Network (HGCN); the corresponding contrastive learning strategy is also introduced to perform representation alignment and prevent negative transfer. Finally, HCLDR can better maintain the balance between diversity and accuracy in the recommender systems, as shown by the experimental results of the study on the Epinions dataset.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 293-304"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145578889","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Infrared spectral imaging-based image recognition for motion detection 基于红外光谱成像的图像识别运动检测

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-01-13 DOI: 10.1016/j.ijin.2025.01.001

Yong Li

{"title":"Infrared spectral imaging-based image recognition for motion detection","authors":"Yong Li","doi":"10.1016/j.ijin.2025.01.001","DOIUrl":"10.1016/j.ijin.2025.01.001","url":null,"abstract":"<div><div>The current infrared imaging recognition methods are inadequate for real-time performance and accuracy for moving objects. Furthermore, they are subject to several constraints, which makes it challenging to recognize stationary and occluded objects. Experts have conducted comprehensive research on infrared imaging, including the development of contour-based infrared motion video image acquisition, the introduction of novel infrared image generation models that align with infrared imaging principles, and the formulation of innovative methods for the joint classification of spatial-spectral and hyper-spectral images. However, none of these advancements have been implemented for enhancement. In order to improve the infrared motion target detection technology, research on image recognition technology based on infrared spectral imaging, the establishment of infrared radiation characteristics model converted image, and combined with the local binary mode for motion target feature extraction, the construction of the background model, applied to the motion detection in the recognition of motion targets. The results demonstrated that the combination effect of local binary pattern feature extraction and analysis of feature vectors increased in accuracy and detection rate with the number of images. Compared to other algorithms, the research algorithm demonstrated a superior signal-to-noise ratio and gain amplitude. The unmanned aerial vehicle signal-to-noise ratio was 13.487, with a gain amplitude of 2.214, while the civil aviation aircraft signal-to-noise ratio was 6.369, with a gain amplitude of 1.792. Therefore, using infrared image feature vectors for image recognition is more effective in motion detection, providing valuable insights for improving the recognition and detection performance of infrared detection technology.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 14-26"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143943163","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Enhancing Time-Sensitive Networking resilience through SDN-based automated failover process 通过基于sdn的自动故障转移过程增强对时间敏感的网络弹性

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-06-04 DOI: 10.1016/j.ijin.2025.05.001

Asha G. Hagargund , Asha K. , Neelavar Shekhar Vittal Shet , Muralidhar Kulkarni

{"title":"Enhancing Time-Sensitive Networking resilience through SDN-based automated failover process","authors":"Asha G. Hagargund , Asha K. , Neelavar Shekhar Vittal Shet , Muralidhar Kulkarni","doi":"10.1016/j.ijin.2025.05.001","DOIUrl":"10.1016/j.ijin.2025.05.001","url":null,"abstract":"<div><div>In domains such as industrial automation, tactile networking, and invehicle communication, stringent requirements for bounded latency and minimal packet loss are paramount to ensure the reliability and efficiency of Time-Sensitive applications. The Time-Sensitive Networking (TSN) aims to cater to these requirements. The architecture of TSN involves heterogeneous data with mixed traffic classes. To ensure the continuous availability of the TSN network, the required failover process for TSN devices must be in place. In this paper, the novel algorithm TSN Device Failover Design (TDFD) for automatic failover configuration of edge switch is proposed and validated using Linux based Open Source tools. Also, the Software Defined Networking (SDN) infrastructure is being employed to enhance the operational efficiency of distributed TSN. For the bounded latency as proposed under the IEEE 802.1Qbv, this work utilizes TAPRIO (Time Aware Priority) queuing discipline. Additionally, the impact of failover on TSN traffic is analyzed by measuring the latency. The experiment result shows that the TSN packets were sent to the destination with a delay of 6 to 13 microseconds before failover. During the failover process, there were no packet transmissions for about 160 ms due to the transition from the old TSN path to the new TSN path due to switch failure. After this, the packets are transmitted to the destination with zero loss. The time taken for the CUC to calculate the new flows and push the new flows to the backup switch is 160 ms.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 176-184"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144922058","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

YOLOv11-RAH: A recurrent attention-enhanced edge intelligence network for UAV-based power transmission line insulator inspection YOLOv11-RAH：用于基于无人机的输电线路绝缘子检测的周期性注意力增强边缘智能网络

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-11-17 DOI: 10.1016/j.ijin.2025.11.008

Yu Liang , Lexin Yang , Shuyi Sun , Zelong Li , Yuyang Shi , Ziteng Zhang , Hong Zhang , Zhiwei Li , Li Zhou , Zhaoyang Zhang , Yihui Shi

{"title":"YOLOv11-RAH: A recurrent attention-enhanced edge intelligence network for UAV-based power transmission line insulator inspection","authors":"Yu Liang , Lexin Yang , Shuyi Sun , Zelong Li , Yuyang Shi , Ziteng Zhang , Hong Zhang , Zhiwei Li , Li Zhou , Zhaoyang Zhang , Yihui Shi","doi":"10.1016/j.ijin.2025.11.008","DOIUrl":"10.1016/j.ijin.2025.11.008","url":null,"abstract":"<div><div>Reliable inspection of power transmission lines is essential for smart grid maintenance, yet manual surveys are slow, costly, and risky, and current detectors still struggle with small insulator defects hidden in complex outdoor scenes. This paper introduces YOLOv11-RAH, a lightweight network framework designed for edge deployment in intelligent inspection systems that pairs the YOLOv11 backbone with a recurrent attention head. The head iteratively refines multi-scale features through depthwise gated convolutions and halts once convergence is reached, strengthening localization without noticeable latency. A four stage augmentation scheme, which are geometric, photometric, Mosaic, and random crop transformations, further improves robustness to viewpoint, illumination, and scale variations. For evaluation, a consolidated insulator corpus is assembled from two cleaned public benchmarks and a newly private UAV dataset, covering diverse outdoor conditions and defect categories. Extensive experiments with other detection algorithms show the consistent accuracy gains of the proposed network improvement, and ablation studies further confirm the benefits of the proposed attention module and augmentation pipeline. In particular, YOLOv11-RAH achieves roughly 5 percentage mAP improvement over YOLOv11n base model on both the consolidated public datasets PTL-AI Furnas and CPLID, and the private UAV dataset AIDD, underscoring its transferability across data sources. These findings suggest that iterative attention combined with targeted data augmentation offers an efficient and reliable solution for automatic insulator condition detection in intelligent inspection networks.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 244-252"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145579025","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Teacher Probability Reconstruction based knowledge distillation within intelligent network compression 基于教师概率重构的智能网络压缩知识蒸馏

International Journal of Intelligent Networks Pub Date : 2025-01-01 Epub Date: 2025-02-22 DOI: 10.1016/j.ijin.2025.02.001

Han Chen , Xuyang Teng , Jiajie Su , Chunhao Li , Chang Hu , Meng Han

{"title":"Teacher Probability Reconstruction based knowledge distillation within intelligent network compression","authors":"Han Chen , Xuyang Teng , Jiajie Su , Chunhao Li , Chang Hu , Meng Han","doi":"10.1016/j.ijin.2025.02.001","DOIUrl":"10.1016/j.ijin.2025.02.001","url":null,"abstract":"<div><div>In the optimization of intelligent network architecture, limited resources at each node, including edge computing devices, have posed challenges for deploying large models in performance-demanding scenarios. Knowledge distillation serves as a model compression method that extracts knowledge from a large-scale teacher model and transfers it to a more lightweight student model. Previous knowledge distillation methods mainly focus on the intermediate layers of the network. However, due to privacy protection regulations that limit data sharing and access as well as computational efficiency requirements in practical scenarios, feature-based distillation encounters challenges in practical applications. We start with logit-based distillation to address these issues, enabling students to learn more representative knowledge from the teacher’s output probability distribution. Due to the structural limitations of the teacher network such as insufficient depth or width, and potential issues in the training data like noise and imbalance, the output probability distribution contains many errors. Therefore, we propose a knowledge distillation method that improves the student by correcting errors in the teacher model. Nevertheless, teacher’s errors not only bring mistakes to students but also give students greater subjectivity, enabling them to break free from the limitations of the teacher. We also retain the teacher’s thinking to prevent students from becoming biased on the remaining (non-target) categories while correcting teacher errors for students. Extensive experiments demonstrate that our method achieves competitive performance on multiple benchmarks without extra parameters.</div></div>","PeriodicalId":100702,"journal":{"name":"International Journal of Intelligent Networks","volume":"6 ","pages":"Pages 47-56"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144070840","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0