检索结果-内蒙古大学图书馆

LIFT+: Lightweight Fine-Tuning for Long-Tail Learning

学校读者我要写书评

暂无评论

arXiv 2025年

作者： Shi, Jiang-Xin Wei, Tong Li, Yu-Feng National Key Laboratory for Novel Software Technology Nanjing University Nanjing210023 China School of Artificial Intelligence Nanjing University Nanjing210023 China School of Computer Science and Engineering Southeast University Nanjing210096 China Key Laboratory of Computer Network and Information Integration Southeast University Ministry of Education China

The fine-tuning paradigm has emerged as a prominent approach for addressing long-tail learning tasks in the era of foundation models. However, the impact of fine-tuning strategies on long-tail learning performance remains unexplored. In this work, we disclose that existing paradigms exhibit a profound misuse of fine-tuning methods, leaving significant room for improvement in both efficiency and accuracy. Specifically, we reveal that heavy fine-tuning (fine-tuning a large proportion of model parameters) can lead to non-negligible performance deterioration on tail classes, whereas lightweight fine-tuning demonstrates superior effectiveness. Through comprehensive theoretical and empirical validation, we identify this phenomenon as stemming from inconsistent class conditional distributions induced by heavy fine-tuning. Building on this insight, we propose LIFT+, an innovative lightweight fine-tuning framework to optimize consistent class conditions. Furthermore, LIFT+ incorporates semantic-aware initialization, minimalist data augmentation, and test-time ensembling to enhance adaptation and generalization of foundation models. Our framework provides an efficient and accurate pipeline that facilitates fast convergence and model compactness. Extensive experiments demonstrate that LIFT+ significantly reduces both training epochs (from ∼100 to ≤15) and learned parameters (less than 1%), while surpassing state-of-the-art approaches by a considerable margin. The source code is available at https://***/shijxcs/LIFT-plus. Copyright © 2025, The Authors. All rights reserved.

关键词：

Tigc-Net: Transformer-Improved Graph Convolution network for Spatio-Temporal Prediction

学校读者我要写书评

暂无评论

SSRN

SSRN 2022年

作者： Chen, Kai Yang, Chunfeng Zhou, Zhengyuan Liu, Yao Ji, Tianjiao Sun, Weiya Chen, Yang School of Cyber Science and Engineering Southeast University Nanjing210096 China Key Laboratory of Computer Network and Information Integration Southeast University Ministry of Education Nanjing210096 China The College of Software Engineering Southeast University Nanjing210096 China Laboratory of Image Science and Technology The School of Computer Science and Engineering Southeast University Nanjing210096 China Jiangsu Key Laboratory of Molecular and Functional Imaging Department of Radiology Zhongda Hospital Southeast University Nanjing210009 China Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing School of Computer Science and Engineering Southeast University Nanjing210096 China NHC Key Laboratory of Medical Virology and Viral Diseases National Institute for Viral Disease Control and Prevention Chinese Center for Disease Control and Prevention Beijing China Beijing Institute of Tracking and Communication Technology Beijing100094 China

Modeling spatio-temporal sequences is an important topic yet challenging for existing neural networks. Most of the current spatio-temporal sequence prediction methods usually capture features separately in temporal and spatial dimensions or employ multiple mutually independent local spatio-temporal graphs to represent a spatio-temporal sequence. The first kind of method mentioned above is difficult to mine the complex spatio-temporal correlations, while the other is limited for the accuracy of long-term predictions. To handle these issues, this paper proposes a Transformer-Improved Graph Convolution network for spatio-temporal prediction. Specifically, the temporal location encoding method is exploited to derive the spatio-temporal characteristics of the sequence utilizing a spatio-temporal feature fusion network. In addition, a spatio-temporal attention network is developed to enhance the spatio-temporal correlation of the sequence, and the dynamic spatial features of sequence are further extracted through the adaptive graph convolution network. A private dataset and a public dataset are employed to demonstrate the performance of the proposed TIGC-Net. The qualitative and quantitative results show that the proposed TIGC-Net can extract dynamic spatiotemporal properties more effectively, enhance the spatio-temporal correlation of sequences and improve the prediction accuracy compared with four stateof-the-art. © 2022, The Authors. All rights reserved.

关键词： Convolution

Variational Gaussian Topic Model with Invertible Neural Projections

学校读者我要写书评

暂无评论

arXiv 2021年

作者： Wang, Rui Zhou, Deyu Xiong, Yuxuan Huang, Haiping School of Computer Science Nanjing University of Posts and Telecommunications China Key Laboratory of Computer Network and Information Integration School of Computer Science and Engineering Ministry of Education Southeast University China Jiangsu High Technology Research Key Laboratory for Wireless Sensor Networks Nanjing China

Neural topic models have triggered a surge of interest in extracting topics from text automatically since they avoid the sophisticated derivations in conventional topic models. However, scarce neural topic models incorporate the word relatedness information captured in word embedding into the modeling process. To address this issue, we propose a novel topic modeling approach, called Variational Gaussian Topic Model (VaGTM). Based on the variational auto-encoder, the proposed VaGTM models each topic with a multivariate Gaussian in decoder to incorporate word relatedness. Furthermore, to address the limitation that pre-trained word embeddings of topicassociated words do not follow a multivariate Gaussian, Variational Gaussian Topic Model with Invertible neural Projections (VaGTM-IP) is extended from VaGTM. Three benchmark text corpora are used in experiments to verify the effectiveness of VaGTM and VaGTMIP. The experimental results show that VaGTM and VaGTM-IP outperform several competitive baselines and obtain more coherent topics. Copyright © 2021, The Authors. All rights reserved.

关键词： Gaussian distribution

Detecting Cryptojacking Traffic Based on network Behavior Features

学校读者我要写书评

暂无评论

Detecting Cryptojacking Traffic Based on Network Behavior Fe...

2021 IEEE Global Communications Conference (GLOBECOM)

作者： Xiaoyan Hu Zhuozhuo Shu Xiaoyi Song Guang Cheng Jian Gong School of Cyber Science & Engineering Southeast University Nanjing China Key Laboratory of Computer Network and Information Integration Southeast University Ministry of Education China Research Base of International Cyberspace Governance (Southeast University) Nanjing China Purple Mountain Laboratories for Network and Communication Security Nanjing China

ISBN: (纸本)9781728181059

Bitcoin and other digital cryptocurrencies have de-veloped rapidly in recent years. To reduce hardware and power costs, many criminals use the botnet to infect other hosts to mine cryptocurrency for themselves, which has led to the proliferation of mining botnets and is referred to as cryptojacking. At present, the mechanisms specific to cryptojacking detection include host-based, Deep Packet Inspection (DPI) based, and dynamic network characteristics based. Host-based detection requires detection installation and running at each host, and the other two are heavyweight. Besides, DPI-based detection is a breach of privacy and loses efficacy if encountering encrypted traffic. This paper de-signs a lightweight cryptojacking traffic detection method based on network behavior features for an ISP, without referring to the payload of network traffic. We set up an environment to collect cryptojacking traffic and conduct a cryptojacking traffic study to obtain its discriminative network traffic features extracted from only the first four packets in a flow. Our experimental study suggests that the machine learning classifier, random forest, based on the extracted discriminative network traffic features can accurately and efficiently detect cryptojacking traffic.

关键词： Privacy Botnet Telecommunication traffic Inspection Feature extraction Hardware Cryptography

Finding the most reliable maximum flow in transport network 2nd

学校读者我要写书评

暂无评论

Finding the most reliable maximum flow in transport network

2nd International Cognitive Cities Conference, IC3 2019

作者： Wang, Jie Cai, Wei Zhou, Sihai Liu, Yundi Liao, Weicheng Zhang, Baili School of Computer Science and Engineering Southeast University Nanjing211189 China Key Laboratory of Computer Network and Information Integration of Ministry of Education Nanjing China Research Center for Judicial Big Data Supreme Count of China Nanjing211189 China

ISBN: (纸本)9789811561122

This paper intends to solve the most reliable maximum flow problem (MRMF) on transport network. A subgraph path division algorithm (SPDA) is proposed to get the most reliable maximum flow distribution, which avoid the negative impact of the number of simple paths and its bottleneck capacity. SPDA divides the sub-graph space of a transport network into a set of disjoint closed intervals, which satisfies the maximum flow constraints. Among the lower bounds of all the intervals, the one with discovered probability has proven to be the most reliable maximum. Finally, experimental results reveal the effectiveness and efficiency of the proposed algorithm. © Springer Nature Singapore Pte Ltd 2020.

关键词： Flow graphs

BRFL: A Blockchain-based Byzantine-Robust Federated Learning Model

学校读者我要写书评

暂无评论

arXiv 2023年

作者： Li, Yang Xia, Chunhe Li, Chang Wang, Tianbo The Key Laboratory of Beijing Network Technology Beihang University Beijing100191 China The Guangxi Collaborative Innovation Center of Multi-Source Information Integration and Intelligent Processing Guangxi Normal University Guilin541004 China The School of Computer Science and Technology Zhengzhou University of Light Industry Zhengzhou450000 China The School of Cyber Science and Technology Beihang University Beijing100191 China The Shanghai Key Laboratory of Computer Software Evaluating and Testing Shanghai201112 China

With the increasing importance of machine learning, the privacy and security of training data have become critical. Federated learning, which stores data in distributed nodes and shares only model parameters, has gained significant attention for addressing this concern. However, a challenge arises in federated learning due to the Byzantine Attack Problem, where malicious local models can compromise the global model's performance during aggregation. This article proposes the Blockchainbased Byzantine-Robust Federated Learning (BRLF) model that combines federated learning with blockchain technology. This integration enables traceability of malicious models and provides incentives for locally trained clients. Our approach involves selecting the aggregation node based on Pearson's correlation coefficient, and we perform spectral clustering and calculate the average gradient within each cluster, validating its accuracy using local dataset of the aggregation nodes. Experimental results on public datasets demonstrate the superior byzantine robustness of our secure aggregation algorithm compared to other baseline byzantine robust aggregation methods, and proved our proposed model effectiveness in addressing the resource consumption problem. Copyright © 2023, The Authors. All rights reserved.

关键词： Blockchain

A Dilated Transformer network for Time Series Anomaly Detection

学校读者我要写书评

暂无评论

A Dilated Transformer Network for Time Series Anomaly Detect...

International Conference on Tools for Artificial Intelligence (ICTAI)

作者： Bo Wu Zhenjie Yao Yanhui Tu Yixin Chen Key Laboratory of Computer Network and Information Integration (Southeast University) Ministry of Education Purple Mountain Laboratories Nanjing China Institute of Microelectronics Chinese Academy of Sciences Purple Mountain Laboratories Beijing China Shandong Future Network Research Institute Purple Mountain Laboratories Beijing China Washington University in St.Louis Purple Mountain Laboratories Washington USA

Unsupervised anomaly detection for time series has been an active research area due to its enormous potential for wireless network management. Existing works have made extraordinary progress in time series representation, reconstruction and forecasting. However, long-term temporal patterns prohibit the model from learning reliable dependencies. To this end, we propose a novel approach based on Transformer with dilated convolution for time anomaly detection. Specifically, we provide a dilated convolution module to extract long-term dependence features. Extensive experiments on various public benchmarks demonstrate that our method has achieved the state-of-the-art performance.

关键词： Convolution Wireless networks Time series analysis Benchmark testing Transformers Feature extraction Market research

Learning from noisy labels via dynamic loss thresholding

学校读者我要写书评

暂无评论

arXiv 2021年

作者： Yang, Hao Jin, Youzhi Li, Ziyin Wang, Deng-Bao Miao, Lei Geng, Xin Zhang, Min-Ling School of Computer Science and Engineering Southeast University Nanjing210096 China Key Laboratory of Computer Network and Information Integration Southeast University Ministry of Education China AI Technology Application Department Huawei Technologies Collaborative Innovation Center of Wireless Communications Technology China

Numerous researches have proved that deep neural networks (DNNs) can fit everything in the end even given data with noisy labels, and result in poor generalization performance. However, recent studies suggest that DNNs tend to gradually memorize the data, moving from correct data to mislabeled data. Inspired by this finding, we propose a novel method named Dynamic Loss Thresholding (DLT). During the training process, DLT records the loss value of each sample and calculates dynamic loss thresholds. Specifically, DLT compares the loss value of each sample with the current loss threshold. Samples with smaller losses can be considered as clean samples with higher probability and vice versa. Then, DLT discards the potentially corrupted labels and further leverages supervised learning techniques. Experiments on CIFAR-10/100 and Clothing1M demonstrate substantial improvements over recent state-of-the-art methods. In addition, we investigate two real-world problems for the first time. Firstly, we propose a novel approach to estimate the noise rates of datasets based on the loss difference between the early and late training stages of DNNs. Secondly, we explore the effect of hard samples (which are difficult to be distinguished) on the process of learning from noisy labels. Copyright © 2021, The Authors. All rights reserved.

关键词： Deep neural networks

Joint Beamforming Design for Integrated Sensing and Communication Systems with Hybrid-Colluding Eavesdroppers

学校读者我要写书评

暂无评论

arXiv 2025年

作者： Liu, Meiding Zhou, Zhengchun Shi, Qiao Li, Guyue Liu, Zilong Fan, Pingzhi Lee, Inkyu School of Information Science and Technology Southwest Jiaotong University Chengdu611756 China Key Laboratory of Analytical Mathematics and Applications [Fujian Normal University Ministry of Education China Purple Mountain Laboratories for Network and Communication Security Nanjing210096 China School of Cyber Science and Engineering Southeast University Nanjing210096 China School of Computer Science and Electronics Engineering University of Essex ColchesterCO4 3SQ United Kingdom Key Laboratory of Information Coding and Transmission Southwest Jiaotong University Chengdu611756 China School of Electrical Engineering Korea University Seoul02841 Korea Republic of

In this paper, we consider the physical layer security (PLS) problem for integrated sensing and communication (ISAC) systems in the presence of hybrid-colluding eavesdroppers, where an active eavesdropper (AE) and a passive eavesdropper (PE) collude to intercept the confidential information. To ensure the accuracy of sensing while preventing the eavesdropping, a base station transmits a signal consisting of information symbols and sensing waveform, in which the sensing waveform can be also used as artificial noise to interfere with eavesdroppers. Under this setup, we propose an alternating optimization-based two stage scheme (AO-TSS) for improving the sensing and communication performance. In the first stage, based on the assumptions that the perfect channel state information (CSI) of the AE and statistical CSI of the PE are known, the communication and sensing beamforming problem is formulated with the objective of minimizing the weighted sum of the beampattern matching mean squared error (MSE) and cross-correlation, subject to the secure transmission constraint. To tackle the non-convexity, we propose a semi-definite relaxation (SDR) algorithm and a reduced-complexity zero-forcing (ZF) algorithm. Then, the scenarios are further extended to more general cases with imperfect AE CSI and unknown PE CSI. To further improve the communication performance, the second-stage problem is developed to optimize the secrecy rate threshold under the radar performance constraint. Finally, numerical results demonstrate the superiority of the proposed scheme in terms of sensing and secure communication. © 2025, CC BY.

关键词： Mean square error