Authentication email has already been sent, please check your email box: and activate it as soon as possible.
You can login to My Profile and manage your email alerts.
If you haven’t received the email, please:
|
|
There are 42 papers published in subject: > since this site started. |
Select Subject |
Select/Unselect all | For Selected Papers |
Saved Papers
Please enter a name for this paper to be shown in your personalized Saved Papers list
|
1. The Analysis of Adversarial Robustness in Federated Learning Based on Differential Privacy | |||
Li Qi,Li Li | |||
Computer Science and Technology 26 March 2024 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:Federated learning (FL) allows multiple participants to train models collaboratively by keeping their data local while exchanging updates, however, this also increases the risk of attack. Models in FL are as vulnerable as centrally trained models against adversarial examples. And with frequent data exchanging, the risk of data breaches increases. To enhance the robustness against adversarial examples, this paper introduces the differential privacy into FL, which is also an effective mechanism used for privacy protection.Many popular approaches that involve adversarial training in FL often compromise test accuracy while effect. In this work, a novel method is proposed which applies differential privacy to enhance both adversarial robustness and privacy protection level. Extensive experiments on multiple datasets demonstrate the proposed approach has a better performance compared with other baselines. | |||
TO cite this article:Li Qi,Li Li. The Analysis of Adversarial Robustness in Federated Learning Based on Differential Privacy[OL].[26 March 2024] http://en.paper.edu.cn/en_releasepaper/content/4763096 |
2. End-to-End Virtual Shadow Generation Based on Shadow Detection | |||
Xue Junsheng, Huang Hai | |||
Computer Science and Technology 21 March 2024 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:With the rapid development of the augmented reality field, virtual shadow generation technology has garnered widespread attention. However, traditional methods involve complex computational requirements for environmental modeling and display, and the effectiveness of using fewer 3D parameter estimation methods is suboptimal. To address these issues, this paper proposes an end-to-end virtual shadow generation method based on shadow detection. Firstly, we introduce a shadow detector designed to identify real shadows and their corresponding occlusions in the background context, while also learning the mapping relationship between occlusions and shadows in the current scene. Secondly, we devise a virtual shadow generator. Utilizing the mapping relationship obtained from the detector as guidance information, it is encoded into the generator's input. Through feature extraction, encoding, and decoding of the input information, we ultimately obtain the virtual shadow image. Experimental results demonstrate the exceptional performance of the proposed virtual shadow generation method. In comparison to existing methods for direct virtual shadow generation, our approach significantly enhances the harmony and realism of the synthesized images. | |||
TO cite this article:Xue Junsheng, Huang Hai. End-to-End Virtual Shadow Generation Based on Shadow Detection[OL].[21 March 2024] http://en.paper.edu.cn/en_releasepaper/content/4762999 |
3. Causal-Proto SSRL: Learning Dynamic-Necessary State Variables for Multi-Environment Reinforcement Learning | |||
Zhang Meng,Zhang Chunhong,Hu Zheng,Zhuang Benhui | |||
Computer Science and Technology 19 March 2024 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:The ability to learn directly from high-dimensional observations such as pixels allows reinforcement learning(RL) to achieve more widely applications. However, high-dimensional observations contain entangled task-relevant and task-irrelevant informations, as well as informations related to actions but unnecessary, which leads to non-essential dependencies and thus affects the generalization and robustness of the reinforcement learning. In order to learn abstract state representations from high-dimensional observations which generalized acorss multiple tasks and robust in environments with different task-unnecessary informations, this paper formulates the POMDP as Partially Observable Temporal Causal Dynamic Models (POTCDMs) and proposes a self-supervised RL with causal representation learning, Causal-Proto RL. This method seperates encoded observations into dynamic-necessary and dynamic-unnecessary state variables where only dynamic-necessary state variables are fed into RL by predicting the causal relationships simultaneously. This method is pretrained in absence of specific task rewards with an intrinsic rewards fo curiosity of causal relationships and are implemented in multiple difficult downstream tasks. This paper evaluate the algorithm in DeepMind Control Suit. This algorithm performs as well as other SOTA slef-supervised RL on a series of downstream tasks in environents as same as pretraining, and demonstrates generalization and robustness in downstream task environments different from pretraining. | |||
TO cite this article:Zhang Meng,Zhang Chunhong,Hu Zheng, et al. Causal-Proto SSRL: Learning Dynamic-Necessary State Variables for Multi-Environment Reinforcement Learning[OL].[19 March 2024] http://en.paper.edu.cn/en_releasepaper/content/4762946 |
4. Research on Option-Critic algorithm based Representation Erasure | |||
Meng JunWei,Hu Zheng | |||
Computer Science and Technology 06 February 2024 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:The Option-Critic (OC) framework can extract transferrable abstract knowledge without requiring any environment-specific prior knowledge, learning options (a form of temporal abstract policy) end-to-end. However, the OC framework exhibits lower data efficiency in transfer tasks. During the learning process, each option considers the entire task's state space, thereby increasing the scale of policy space search. This paper proposes an Option Learning algorithm based on Representation Erasure, which introduces the Representation Erasure method to clearly quantify the influence of each dimension on high-level and low-level policy learning. It identifies and erases dimensions that significantly interfere with training, effectively reducing the scale of policy space search. Through theoretical derivation and experimental validation, this paper demonstrates the effectiveness of the Representation Erasure-based Option Learning algorithm. | |||
TO cite this article:Meng JunWei,Hu Zheng. Research on Option-Critic algorithm based Representation Erasure[OL].[ 6 February 2024] http://en.paper.edu.cn/en_releasepaper/content/4762077 |
5. End-to-end 3D Human Pose Estimation using Dual Decoders | |||
WANG Zhang,SONG Mei,JIN Lei | |||
Computer Science and Technology 02 February 2024 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:Existing methods for 3D human pose estimation mainly divide the task into two stages. The first stage identifies the 2D coordinates of the human joints in the input image, namely the 2D human joint coordinates. The second stage uses the results from the first stage as input to recover the depth information of human joints from the 2D human joint coordinates to achieve 3D human pose estimation. However, the recognition accuracy of the two-stage method relies heavily on the results of the first stage and includes too many redundant processing steps, which reduces the inference efficiency of the network. To address these issues, we propose the EDD, a fully End-to-end 3D human pose estimation method based on transformer architecture with Dual Decoders. By learning multiple human poses, the model can directly infer all 3D human poses in the image using a pose decoder, and then further optimize the recognition result using a joint decoder based on the kinematic relations between joints. With the attention mechanism, this method can adaptively focus on the most relevant features to the target joint, effectively overcoming the feature misalignment problem in the human pose estimation task and greatly improving the model performance. Any complex post-processing step, such as non-maximum suppression, is eliminated, further improving the efficiency of the model. The results show that this method achieves an accuracy of 87.4\% on the MuPoTS-3D dataset, significantly improving the accuracy of the end-to-end 3D human pose estimation method based on mixed training. | |||
TO cite this article:WANG Zhang,SONG Mei,JIN Lei. End-to-end 3D Human Pose Estimation using Dual Decoders[OL].[ 2 February 2024] http://en.paper.edu.cn/en_releasepaper/content/4761949 |
6. Research on Deep Learning Stock Selection Method Based on Trend Decomposition Algorithm | |||
Wu Cheng-Hui,Zhang Hong-Jian | |||
Computer Science and Technology 20 June 2023 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:Some studies using machine learning and deep learning models did not fully explore other indicator features that affect stock price trends, resulting in significant prediction errors and difficulty in directly applying the prediction results to obtain excess returns. Therefore, it is necessary to establish an efficient deep learning model based on stock characteristics and fully explore the factors that affect stock price trends to improve prediction accuracy.In view of the shortcomings of traditional stock trend prediction, this paper constructs LSTM-BP neural network based on Tianniuxu optimization for stock price prediction, and constructs PCA-BP neural network model based on particle swarm optimization based on multi angle correlation analysis for trend prediction. Taking into account both price prediction and trend prediction dimensions, construct a deep quantitative stock selection method. In this process, a stock trend decomposition judgment algorithm is proposed to identify and classify the stock data, improve the data utilization, build a new data set, and use the particle swarm optimization to adjust and optimize the parameters.The results of strategy backtesting show that compared to other deep learning strategies and benchmark strategies, the stock selection method and strategy proposed in this article have significantly improved in terms of victory and return. This indicates that the model has good predictive ability and application potential, and can provide effective decision support for stock investment. | |||
TO cite this article:Wu Cheng-Hui,Zhang Hong-Jian. Research on Deep Learning Stock Selection Method Based on Trend Decomposition Algorithm[OL].[20 June 2023] http://en.paper.edu.cn/en_releasepaper/content/4760906 |
7. Federated learning based on hybrid blockchain | |||
FAN Linxuan,LI Lixiang | |||
Computer Science and Technology 16 March 2023 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:With the rapid development of technologies such as machine learning, 5G communication, edge computing, artificial intelligence and blockchain, the field of machine learning has produced some new training methods. Among them, federated learning is a typical representative of distributed machine learning. Compared with traditional machine learning, federated learning can collaborate on model training between organizations without exchanging original data sets, which ensures the security and privacy of organizational data. As a kind of distributed machine learning, federated learning is faced with severe technical challenges in the process of model training: low participation in the training of edge nodes, untrustworthy edge nodes and untraceable training data. To solve the above problems, based on the federated learning theory and the blockchain theory, this paper carries out relevant research on the federated learning algorithm based on hybrid blockchain and the blockchain incentive mechanism. The main research achievements and innovations of this paper are as follows: the existing federated learning algorithm hides training data, which gives the attacker an opportunity to exploit, and the attacker can use this defect to carry out backdoor attacks on model training. In addition, in the training process of federated learning algorithm, the identity of the participating nodes is not authenticated, so that the attacker can pretend the nodes to contribute dirty data, which reduces the accuracy of model training. Therefore, this paper proposes a federated learning algorithm based on hybrid blockchain, which mainly adopts consortium blockchain to authenticate and manage the identity of nodes participating in training. Meanwhile, public blockchain is used to store training parameters to achieve traceability of training data. In addition, the introduction of blockchain architecture enables federated learning to be further decentralized. The results of simulation experiments show that the proposed scheme has advantages in robustness and accuracy of model training under the same training task. | |||
TO cite this article:FAN Linxuan,LI Lixiang. Federated learning based on hybrid blockchain[OL].[16 March 2023] http://en.paper.edu.cn/en_releasepaper/content/4759526 |
8. KDText: A Lightweight Scene Text Detector with Decoupling-Based Knowledge Distillation | |||
Lei Siyue,Yan DanFeng | |||
Computer Science and Technology 28 February 2023 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:In recent years, scene text detection task has made great progress. However, most of works are devoted to improving the performance of detector and pay little attention to the practical applications. In this paper, we propose to exploit a mask branch to detect arbitrary-shaped text accurately, and compress the model to reduce computation and storage costs, achieving a balance between speed and accuracy. Specifically, we distill intermediate features and text proposal classification to transfer dark knowledge to student text detector. In the distillation process, we treat textual and Background features differently and decouple positive and negative text proposals. Experimental results on the ICDAR 2015 and ICDAR 2017 MLT datasets demonstrate the superiority of our lightweight scene text detector. | |||
TO cite this article:Lei Siyue,Yan DanFeng. KDText: A Lightweight Scene Text Detector with Decoupling-Based Knowledge Distillation[OL].[28 February 2023] http://en.paper.edu.cn/en_releasepaper/content/4759241 |
9. RKDE: Data Compression for Time-Series Data based on Kernel Density Estimation in Reservoir Algorithm . | |||
CAO Yanyun,XU Peng | |||
Computer Science and Technology 15 February 2023 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:Edge-cloud collaborative anomaly detection has become the most important anomaly detection architecture. However, only in the most ideal state can the central cloud platform be fully trained with sufficient data. In the case of limited communication, we have to consider reducing the use of communication resources, but still maintain a high accuracy rate of anomaly detection. In this context, RKDE, a reservoir sampling algorithm based on kernel density estimation, is proposed to reduce the amount of data uploaded to the cloud by the edge end. By improving the probability of abnormal data being sampled, the compression pool of upload is constructed, the redundant process in gradient exchange is reduced, and the sampling process is timely fed back and adjusted according to the abnormal detection results in the cloud. At the same time, RKDE is compared with several sampling algorithms to demonstrate its performance advantages. | |||
TO cite this article:CAO Yanyun,XU Peng. RKDE: Data Compression for Time-Series Data based on Kernel Density Estimation in Reservoir Algorithm .[OL].[15 February 2023] http://en.paper.edu.cn/en_releasepaper/content/4759073 |
10. An Image Steganalysis Model Based on Local Difference Analysis | |||
Han Ming,Jin ZhuJun,Yang Yu | |||
Computer Science and Technology 01 February 2023 | |||
Show/Hide Abstract | Cite this paper︱Full-text: PDF (0 B) | |||
Abstract:Current image steganalysis models usually extract global features to distinguish between natural image and steganographic image. However, the global features that represent the differences between images are closely related to image properties, such as embedding payload, which will mostly lead to severe performance deterioration if there is a mismatch between training and detecting sources. On the contrary, the local feature changes of the moving window in an image are mainly affected by data hiding, but not by the image properties. Therefore, a detection method based on local difference analysis is proposed in this paper. By analyzing the smoothness changes of local moving blocks, the impact of payload mismatch on steganalysis accuracy is reduced. In addition, this paper also introduces long short-term memory network technique into image steganalysis, and proposes a new detection model called ILDA-Net (Image Local Difference Analysis Net), which analyzes the changes in the local residual sequence to achieve steganalysis. Experiments on LSB and WOW steganography algorithms show that ILDA-Net can effectively reduce the impact of payload mismatch on network detection performance. | |||
TO cite this article:Han Ming,Jin ZhuJun,Yang Yu. An Image Steganalysis Model Based on Local Difference Analysis[OL].[ 1 February 2023] http://en.paper.edu.cn/en_releasepaper/content/4758921 |
Select/Unselect all | For Selected Papers |
Saved Papers
Please enter a name for this paper to be shown in your personalized Saved Papers list
|
About Sciencepaper Online | Privacy Policy | Terms & Conditions | Contact Us
© 2003-2012 Sciencepaper Online. unless otherwise stated