8 results on '"Li, Yuancheng"'
Search Results
2. Clean-label attack based on negative afterimage on neural networks
- Author
-
Zang, Liguang and Li, Yuancheng
- Published
- 2024
- Full Text
- View/download PDF
3. Virtual network function deployment algorithm based on graph convolution deep reinforcement learning.
- Author
-
Qiu, Rixuan, Bao, Jiawen, Li, Yuancheng, Zhou, Xin, Liang, Liang, Tian, Hui, Zeng, Yanting, and Shi, Jie
- Subjects
REINFORCEMENT learning ,VIRTUAL networks ,MACHINE learning ,GRAPH algorithms ,LINEAR programming ,INTEGER programming ,QUALITY of service - Abstract
In Network Function virtualization (NFV), network functions are virtualized as Virtual Network functions (VNFs). A network service consists of a set of VNFs. One of the major challenges in implementing this paradigm is allocating optimal resources to VNFs. Most existing work assumes that services are represented as service functional chains (SFC), which are chains. However, for more complex and diversified network services, a more appropriate representation is Virtual Network Function Forwarding Graph (VNF-FG), namely directed acyclic Graph (DAGs). Previous works failed to take advantage of this special graph structure, which makes them unsuitable for complex and diverse network service scenarios. Aiming at the problem of virtual network function Forwarding Graph mapping (VNF-FGE) represented by DAG, this paper proposes a virtual network function (VNF) deployment algorithm based on graph convolution network (GCN) and deep reinforcement learning (DRL), called GDRL-VNFP. First, we describe the VNF-FGE problem as an integer linear programming (ILP) problem and the virtual network service request (NSR) as a DAG. Secondly, in order to overcome the challenges brought about by the different sizes and dynamic arrival of the DAG representation of NSR, an efficient algorithm based on GCN and DRL is proposed, which can generate deployment solutions in real time under the premise of meeting the quality of service and minimize the total cost of deployment. We use GCN to extract features from the physical network topology and VNF-FG represented by DAG, and in addition, and construct a sequence-to-sequence model based on the attention mechanism for the output mapping scheme. Finally, we trained the model using a policy-based gradient-based reinforcement learning algorithm that could quickly find a near-optimal solution for the problem instance. Simulation results show that our GDRL-VNFP outperforms state-of-the-art solutions in terms of deployment cost, service request reception rate and runtime. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF
4. An Efficient Method of Tooth Segmentation Under Massive Medical Data.
- Author
-
Ma, Tian, Yang, Yizhou, Li, Yun, Li, Zhanli, and Li, Yuancheng
- Subjects
TEETH ,HUMAN-computer interaction - Abstract
To accurately and efficiently complete tooth segmentation from a large amount of oral medical data, the burden of doctors should be reduced. An automatic seed picking method based on 2D projection of the occlusal plane was proposed. First, the authors establish a two-dimensional seed data set for tooth segmentation. Then, this article built a prediction network of teeth seeds based on YOLOv4 to realize the prediction of teeth position as well as the recognition of teeth categories. Finally, according to the statistical optimal seed position, the two-dimensional seeds are calculated and mapped back to the three-dimensional space by the reverse projection transformation method to realize the final picking up of the three-dimensional seeds. Furthermore, combined with the previous work of division line detection, the automatic segmentation of the 3D dental model was realized. The experimental results show that the proposed method has high accuracy and real-time performance, which significantly reduces the burden of human-computer interaction in dental model segmentation. [ABSTRACT FROM AUTHOR]
- Published
- 2022
- Full Text
- View/download PDF
5. Poisonous Label Attack: Black-Box Data Poisoning Attack with Enhanced Conditional DCGAN.
- Author
-
Liu, Haiqing, Li, Daoxing, and Li, Yuancheng
- Subjects
POISONS ,CONVOLUTIONAL neural networks ,POISONING ,GENERATIVE adversarial networks ,MACHINE learning - Abstract
Data poisoning is identified as a security threat for machine learning models. This paper explores the poisoning attack against the convolutional neural network under black-box conditions. The proposed attack is "black-box," which means the attacker has no knowledge about the targeted model's structure and parameters when attacking the model, and it uses "poisonous-labels" images, fake images with crafted wrong labels, as poisons. We present a method for generating "poisonous-label" images that use Enhanced Conditional DCGAN (EC-DCGAN) to synthesizes fake images and uses asymmetric poisoning vectors to mislabel them. We evaluate our method by generating "poisonous-label" images from MNIST and FashionMNIST datasets and using them to manipulate image classifiers. Our experiments demonstrate that, similarly to white box data poisoning attacks, the poisonous label attack can also dramatically increase the classification error. [ABSTRACT FROM AUTHOR]
- Published
- 2021
- Full Text
- View/download PDF
6. Denial of service attack and defense method on load frequency control system.
- Author
-
Li, Yuancheng, Zhang, Pan, and Ma, Longqiang
- Subjects
- *
DENIAL of service attacks , *PHASOR measurement , *MACHINE learning , *ELECTRIC power distribution grids , *CONTROL rooms , *DEEP learning - Abstract
• This paper proposed a novel method to defense DOS attack on LFC in smart grid. • The LFC model of two area power system with DOS attack is analyzed. • An improved data prediction algorithm is proposed. • The effectiveness of the proposed method is verified by simulation experiments. The application of the network technology in the power grid makes the Load Frequency Control (LFC) system more vulnerable to various kinds of network attacks. The Denial of Service (DOS) attack can block the data collected by the Phasor measurement unit from being transmitted to the LFC center, thereby affecting the decision of the control center and generation of control signals, and can not adjust the frequency of the power grid timely. Aiming at the DOS attack on LFC, a defense method based on data prediction is proposed. Through the combination of the deep learning algorithm and the Extreme Learning Machine (ELM) algorithm, the Deep auto-encoder Extreme Learning Machine (DAELM) algorithm combines the advantages of the fast speed of the extreme learning machine and the advantages of high accuracy of the deep learning. We can predict and supplement the lost data based on the DAELM algorithm, and ensure the normal operation of the LFC system, thus can prevent DOS attacks. The experiments verified the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]
- Published
- 2019
- Full Text
- View/download PDF
7. Attention-aware deep reinforcement learning for detecting false data injection attacks in smart grids.
- Author
-
Huang, Rong, Li, Yuancheng, and Wang, Xiao
- Subjects
- *
MACHINE learning , *REINFORCEMENT learning , *SELECTIVITY (Psychology) , *FEATURE extraction , *CYBERTERRORISM , *ELECTRONICS engineers - Abstract
• As the attack vectors injected by FDIA in power systems become more and more concealed, a model-free FDIA detection measurement method focusing on state attention is proposed. • Extracting potential features of the state vector rather than distracting from irrelevant details by introducing an attention mechanism to improve the discrimination of state observations in the MDP and to guide agents to make rational decisions. • Attention weights of each measurement point reveal the potential behavioral changes of attack points and attack strength selection when attackers construct attacks. • The proposed method is experimented on IEEE-14 and IEEE-118 bus systems and demonstrates the method's satisfaction in terms of detection accuracy and efficiency. Cyber-attacks have long undermined the security and stable operation of the smart grid. The increasing share of intermittent energy aggravates the uncertainty of the grid environment and makes it easier for attacks to be stealthy. In recent research, attackers can construct highly stealthy false data injection attacks (FDIAs) with as little-known information as possible. Faced with a volatile grid environment and severe cyber threats, we propose a deep reinforcement learning FDIAs detection method. It focuses on state attention to address the problems of state feature extraction in existing reinforcement learning detection methods. The proposed approach adds an attention mechanism to the model-free deep reinforcement learning detection algorithm to extract state features. Selective attention helps us focus on the important parts instead of distracting from irrelevant details, making the states more representative and distinguishable, thus improving the reinforcement learning framework to detect attacks faster and more accurately. The proposed method has experimented on IEEE-14, and IEEE-118 bus systems provided by the Institute of Electrical and Electronics Engineers (IEEE) and demonstrates the method's satisfaction in terms of detection accuracy and efficiency compared with similar methods. © 2017 Elsevier Inc. All rights reserved. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF
8. Curvelet based image compression via core vector machine.
- Author
-
Li, Yuancheng, Wang, Yiliang, Xiao, Rui, and Yang, Qiu
- Subjects
- *
CURVELET transforms , *IMAGE compression , *DIGITAL image processing , *MACHINE learning , *SUPPORT vector machines , *SIGNAL-to-noise ratio - Abstract
Abstract: In this paper, a novel curvelet based digital image compression scheme is proposed. Aiming at achieving high compression ratio, the proposed scheme embeds a representative machine learning method, core vector machine (CVM), in the encoding process of the image compression technique. The core vector machine (CVM) has been introduced as an extremely fast classifier which is demonstrably superior to standard support vector machine (SVM) on very large datasets. In this scheme, we appropriately utilize the characteristic of CVM to reduce huge numbers of curvelet coefficients. Compared with image compression algorithms do not use CVM and methods based on wavelet transform, experimental results show that the compression performance of our method gains much improvement in peak-signal-to-noise-ratio (PSNR) and CPU time. Moreover, the algorithm works fairly well for declining block effect at higher compression ratios. [Copyright &y& Elsevier]
- Published
- 2013
- Full Text
- View/download PDF
Catalog
Discovery Service for Jio Institute Digital Library
For full access to our library's resources, please sign in.