检索结果-内蒙古大学图书馆

Adversarial Attacks on Large Language Model-Based System and Mitigating Strategies: A Case Study on ChatGPT

Security and Communication Networks 2023年第1期2023卷

作者： Liu, Bowen Xiao, Boao Jiang, Xutong Cen, Siyuan He, Xin Dou, Wanchun State Key Laboratory for Novel Software Technology Nanjing University Nanjing China Guangdong Laboratory of Artificial Intelligence and Digital Economy Shenzhen China School of Computer Science & Technology Nanjing University of Posts and Telecommunications Nanjing China College of Big Data and Intelligent Engineering Southwest Forestry University Kunming China

Machine learning algorithms are at the forefront of the development of advanced information systems. The rapid progress in machine learning technology has enabled cutting-edge large language models (LLMs), represented by GPT-3 and ChatGPT, to perform a wide range of NLP tasks with a stunning performance. However, research on adversarial machine learning highlights the need for these intelligent systems to be more robust. Adversarial machine learning aims to evaluate attack and defense mechanisms to prevent the malicious exploitation of these systems. In the case of ChatGPT, adversarial induction prompt can cause the model to generate toxic texts that could pose serious security risks or propagate false information. To address this challenge, we first analyze the effectiveness of inducing attacks on ChatGPT. Then, two effective mitigating mechanisms are proposed. The first is a training-free prefix prompt mechanism to detect and prevent the generation of toxic texts. The second is a RoBERTa-based mechanism that identifies manipulative or misleading input text via external detection models. The availability of this method is demonstrated through experiments. © 2023 Bowen Liu et al.

关键词： Machine learning

来源：评论

学校读者我要写书评

暂无评论

Towards Combating Frequency Simplicity-biased Learning for Domain Generalization

arXiv

引用

arXiv 2024年

作者： He, Xilin Hu, Jingyu Lin, Qinliang Luo, Cheng Xie, Weicheng Song, Siyang Khan, Muhammad Haris Shen, Linlin Computer Vision Institute School of Computer Science & Software Engineering Shenzhen University China Shenzhen Institute of Artificial Intelligence and Robotics for Society China Guangdong Provincial Key Laboratory of Intelligent Information Processing China University of Exeter United Kingdom Mohamed bin Zayed University of Artificial Intelligence United Arab Emirates

Domain generalization methods aim to learn transferable knowledge from source domains that can generalize well to unseen target domains. Recent studies show that neural networks frequently suffer from a simplicity-biased learning behavior which leads to over-reliance on specific frequency sets, namely as frequency shortcuts, instead of semantic information, resulting in poor generalization performance. Despite previous data augmentation techniques successfully enhancing generalization performances, they intend to apply more frequency shortcuts, thereby causing hallucinations of generalization improvement. In this paper, we aim to prevent such learning behavior of applying frequency shortcuts from a data-driven perspective. Given the theoretical justification of models’ biased learning behavior on different spatial frequency components, which is based on the dataset frequency properties, we argue that the learning behavior on various frequency components could be manipulated by changing the dataset statistical structure in the Fourier domain. Intuitively, as frequency shortcuts are hidden in the dominant and highly dependent frequencies of dataset structure, dynamically perturbating the over-reliance frequency components could prevent the application of frequency shortcuts. To this end, we propose two effective data augmentation modules designed to collaboratively and adaptively adjust the frequency characteristic of the dataset, aiming to dynamically influence the learning behavior of the model and ultimately serving as a strategy to mitigate shortcut learning. Code is available at AdvFrequency. © 2024, CC BY.

关键词： Contrastive Learning

来源：评论

学校读者我要写书评

暂无评论

ASRNet: Adversarial Segmentation and Registration Networks for Multispectral Fundus Images

引用

Computer Systems Science & engineering 2021年第3期36卷 537-549页

作者： Yanyun Jiang Yuanjie Zheng Xiaodan Sui Wanzhen Jiao Yunlong He Weikuan Jia School of Information Science and Engineering Shandong Normal UniversityJinan250358China Key Laboratory of Intelligent Computing&Information Security in Universities of Shandong Shandong Provincial Key Laboratory for Novel Distributed Computer Software TechnologyInstitute of Biomedical SciencesShandong Normal UniversityJinan250358China Shandong Provincial Hospital Affiliated to Shandong University Jinan250021China INSA Lyon University of LyonCNRSInsermVilleurbanne69621CedexFrance

Multispectral imaging (MSI) technique is often used to capture imagesof the fundus by illuminating it with different wavelengths of light. However,these images are taken at different points in time such that eyeball movementscan cause misalignment between consecutive images. The multispectral imagesequence reveals important information in the form of retinal and choroidal bloodvessel maps, which can help ophthalmologists to analyze the morphology of theseblood vessels in detail. This in turn can lead to a high diagnostic accuracy of several diseases. In this paper, we propose a novel semi-supervised end-to-end deeplearning framework called “Adversarial Segmentation and Registration Nets”(ASRNet) for the simultaneous estimation of the blood vessel segmentation andthe registration of multispectral images via an adversarial learning process. ASRNet consists of two subnetworks: (i) A segmentation module S that fulfills theblood vessel segmentation task, and (ii) A registration module R that estimatesthe spatial correspondence of an image pair. Based on the segmention-drivenregistration network, we train the segmentation network using a semi-supervisedadversarial learning strategy. Our experimental results show that the proposedASRNet can achieve state-of-the-art accuracy in segmentation and registrationtasks performed with real MSI datasets.

关键词： Deep learning deformable image registration image segmentation multispectral imaging(MSI)

来源：评论

学校读者我要写书评

暂无评论

Driving at Realistic Urban Intersections with Adversarial Reinforcement Learning

Driving at Realistic Urban Intersections with Adversarial Re...

引用

Automated Vehicle Validation Conference (IAVVC), IEEE International

作者： Yangang Ren Letian Tao Jingliang Duan Honglin Li Beiyan Jiang Shengbo Eben Li State Key Laboratory of Intelligent Vehicle Safety Technology Chongqing Changan Automobile CO. Ltd. Chongqing China School of Vehicle and Mobility Tsinghua University Beijing China School of Mechanical Engineering University of Science and Technology Beijing Beijing China Center of Software Engineering Research Dongfeng Motor Corporation Research and Development Institute Wuhan China

ISBN: (数字)9798350354072

ISBN: (纸本)9798350354089

Reinforcement learning (RL) has been widely studied in the decision-making of automated vehicles and demonstrated a promising self-evolution ability. However, these RL-driven methods are usually carried out based on simulation due to safety concerns of the interaction with the real world. This results in the policy well-trained in simulation struggles to remain effective in realistic environments. In this paper, we propose an adversarial training method to improve the practical performance of driving policy at urban roads. The basic idea is to proactively model the discrepancy between simulation and reality as the adversarial noise, and design offline competitive learning to enhance the robustness against environment variation. The driving policy jointly minimizes the path tracking and safety cost, while the adversary generates refined disturbances on the simulation model to increase the collision risk. The smooth adversarial actor-critic algorithm is adopted to update these policies iteratively with a gradient-based manner. Then, the trained policy is deployed on a full-size vehicle to verify its effectiveness at an urban intersection. Experiment results show that the policy can not only control the vehicle smoothly and efficiently, but also deal with the abnormal or unseen behaviors from surrounding traffic actors. Our work provides a feasibility for the application of RL-based methods in real-world driving.

关键词： Training Constraint optimization Costs Roads Noise Decision making Reinforcement learning Games Robustness Safety

来源：评论

学校读者我要写书评

暂无评论

Efficient Test-Time Model Adaptation without Forgetting

arXiv

引用

arXiv 2022年

作者： Niu, Shuaicheng Wu, Jiaxiang Zhang, Yifan Chen, Yaofo Zheng, Shijian Zhao, Peilin Tan, Mingkui School of Software Engineering South China University of Technology China Key Laboratory of Big Data and Intelligent Robot Ministry of Education China Tencent AI Lab China National University of Singapore Singapore Pazhou Laboratory China

Test-time adaptation (TTA) seeks to tackle potential distribution shifts between training and testing data by adapting a given model w.r.t. any testing sample. This task is particularly important for deep models when the test environment changes frequently. Although some recent attempts have been made to handle this task, we still face two practical challenges: 1) existing methods have to perform backward computation for each test sample, resulting in unbearable prediction cost to many applications;2) while existing TTA solutions can significantly improve the test performance on out-of-distribution data, they often suffer from severe performance degradation on in-distribution data after TTA (known as catastrophic forgetting). In this paper, we point out that not all the test samples contribute equally to model adaptation, and high-entropy ones may lead to noisy gradients that could disrupt the model. Motivated by this, we propose an active sample selection criterion to identify reliable and non-redundant samples, on which the model is updated to minimize the entropy loss for test-time adaptation. Furthermore, to alleviate the forgetting issue, we introduce a Fisher regularizer to constrain important model parameters from drastic changes, where the Fisher importance is estimated from test samples with generated pseudo labels. Extensive experiments on CIFAR-10-C, ImageNet-C, and ImageNet-R verify the effectiveness of our proposed method. Code is available at https://***/mr-eggplant/EATA. Copyright © 2022, The Authors. All rights reserved.

关键词： Entropy

来源：评论

学校读者我要写书评

暂无评论

Uncertainty-Aware Modeling for Improving Prediction of Toxicity

Uncertainty-Aware Modeling for Improving Prediction of Toxic...

引用

International Conference on Machine Learning and Cybernetics (ICMLC)

作者： Xinlei Zhou Xizhao Wang Xinpeng Zhou Siwu Tu Ranwang School of Mathematical Sciences Shenzhen University Shenzhen China College of Computer Science and Software Engineering Shenzhen University Shenzhen China Guangdong Provincial Key Laboratory of Intelligent Information Processing Shenzhen University Shenzhen China College of Electronic Information Guangxi Minzu University 530000 China Jiangxi Drug Inspection Center Nanchang China

ISBN: (数字)9798331528041

ISBN: (纸本)9798331528058

Accurately predicting molecular toxicity is challenging in drug discovery and chemical safety assessment. Except for the representation of molecular structure, the inherent class imbalance and uncertainty often complicate the task. In this work, we present UnBGMT (Uncertainty-Aware Balanced Graph Multi-set Transformer), which integrates uncertainty modeling into the Graph Multiset Transformer (GMT) framework to improve the reliability and accuracy of molecular toxicity predictions. Our method employs Monte Carlo dropout (MC dropout) to evaluate uncertainty, which is then used to guide an adaptive loss function that adjusts weights in training. This approach guides the model to focus on the minority and samples with middle-level uncer-tainty, mitigating issues related to class imbalance and improving model stability. We conducted sufficient experiments on the Tox21 dataset, demonstrating the superiority of UnBGMT by comparing performance with baseline models and conducting ablation studies. Our results highlight the benefits of incorporating uncertainty in toxicity prediction, which provides a deeper insight into the im-pact of uncertainty on model performance.

关键词： Adaptation models Uncertainty Toxicology Molecular communication Predictive models Transformers Stability analysis Safety Drug discovery Chemicals

来源：评论

学校读者我要写书评

暂无评论

A Weighted Flat Lattice Transformer-based Knowledge Extraction Architecture for Chinese Named Entity Recognition

A Weighted Flat Lattice Transformer-based Knowledge Extracti...

引用

International Conference on Computer Supported Cooperative Work in Design

作者： Hengwei Zhang Yuejia Wu Jian-Tao Zhou Inner Mongolia University College of Computer Science Hohhot China Ministry of Education National & Local Joint Engineering Research Center of Intelligent Information Processing Technology for Mongolian Engineering Research Center of Ecological Big Data Hohhot China Inner Mongolia Engineering Laboratory for Cloud Computing and Service Software Hohhot China Inner Mongolia Key Laboratory of Social Computing and Data Processing Hohhot China Inner Mongolia Key Laboratory of Discipline Inspection and Supervision Big Data Hohhot China Inner Mongolia Engineering Laboratory for Big Data Analysis Technology Hohhot China

ISBN: (数字)9798350349184

ISBN: (纸本)9798350349191

Named Entity Recognition (NER) is one of the contents of Knowledge Extraction (KE) that transforms data into knowledge representation. However, Chinese NER faces the problem of lacking clear word boundaries that limit the effectiveness of the KE. Although the flat lattice Transformer (FLAT) framework, which converts lattice structure into a flat structure including a set of spans, can effectively improve this problem and obtain advanced results, there still exist the problems of insensitivity to entity importance weights and insufficient feature learning. This paper proposes a weighted flat lattice Transformer architecture for Chinese NER, namely WFLAT. The WFLAT first adds a weight matrix into self-attention calculation, which can obtain finer-grained partitioning of entities to improve experimental performance, and then adopts a multi-layer Transformer encoder with each layer using a multi-head self-attention mechanism. Extensive experimental results on benchmarks demonstrate that our proposed KE model can obtain state-of-the-art performance for the Chinese NER task.

关键词： Representation learning Computational modeling Memory management Lattices Named entity recognition Transforms Knowledge representation

来源：评论

学校读者我要写书评

暂无评论

Delay Analysis of Multi-Priority Computing Tasks in Alibaba Cluster Traces

Delay Analysis of Multi-Priority Computing Tasks in Alibaba ...

引用

IEEE Conference on Computer Communications Workshops, INFOCOM Wksps

作者： Chenyu Gong Mulei Ma Liekang Zeng Yang Yang Xiaohu Ge Liantao Wu IoT Thrust and Research Center for Digital World with Intelligent Things (DOIT) The Hong Kong University of Science and Technology (Guangzhou) Guangzhou China Peng Cheng Laboratory Shenzhen China Terminus Group Beijing China Huazhong University of Science and Technology China Software Engineering Institute East China Normal University Shanghai China

ISBN: (数字)9798350384475

ISBN: (纸本)9798350384482

Future 6G network AI architecture will provide customized services for everyone, so it is necessary to concentrate on user requirements. Most applications adopt delay as a critical indicator, but few studies have explored its distribution's underlying causes in detail. To address this issue, we analyzed some comparable datasets, i.e., Alibaba Cluster Traces for the years 2017 and 2018, aiming to get some insights from the facts. Traces contain both high-priority online tasks and low-priority batch tasks. We observed the distributions of multiple fields and found several insights: (i) Due to the low priority, the planned CPU cores of batch tasks will be limited by the online tasks. In other words, the service delay of batch tasks will be impacted by the high-priority online tasks. (ii) The overall distribution of task delays and planned CPU cores has upper and lower bounds, both of which can be fitted with logarithmic functions. (iii) The future network AI architecture will be multi-tier and resource fine-grained. Because the flexibility of resource allocation for batch tasks will increase as the CPU granularity of the online tasks grows.

关键词： 6G mobile communication Computer architecture Market research Real-time systems Delays Resource management Security

来源：评论

学校读者我要写书评

暂无评论

Dsh-Rsf-Lite Approach: Towards Lightweight Pneumonia Classification Network with Few-Shot Augmentation on Chest X-Ray Images

SSRN

引用

SSRN 2024年

作者： Chen, Hongbin Zhu, Zhou Wang, Wei Yang, Chunfeng Li, Jie Liu, Chengyu Kiong, Loo Chu Wang, Zhelong Li, Jianqing Xiang, Wentao The Engineering Research Center of Intelligent Theranostics Technology and Instruments Ministry of Education School of Biomedical Engineering and Informatics Nanjing Medical University Nanjing211166 China The School of Software Engineering Jinling Institute of Technology Nanjing211169 China Nanjing China The State Key Laboratory of Digital Medical Engineering School of Instrument Science and Engineering Southeast University Nanjing China Kuala Lumpur50603 Malaysia The School of Control Science and Engineering Dalian University of Technology Dalian116024 China

Pneumonia classification using chest X-ray images has significant clinical diagnosis values with its rapid image acquisition, low radiation dose and rich volume ***, they may suffer from class imbalance and redundant characteristics. To address these issues, the DSH-RSF-Lite approach is proposed with three core parts: 1) DSHNet is a few-shot augmentation approach that integrates spectral normalization into deep convolutional generative adversarial network and employs hinge adversarial loss to augment chest X-ray images for balancing positive and negative samples, 2) RSFNet is a residual network that incorporates spatial attention blocks and early-late feature fusion, enhancing the model focus on key regions within chest X-ray images meanwhile enriching overall information representation, 3) RSF-Lite is a lightweight model that employs geometric median-based pruning strategy to eliminate redundant filters in RSFNet, significantly reducing model parameters and computational complexity. Experimental results on the chest X-ray 2017 dataset demonstrate that: 1) our proposed DSHNet achieves Fréchet Inception Distance of 36.92 and Kernel Inception Distance of 0.03 for generating chest X-ray images, outperforming three classical GAN-based approaches, 2) our proposed DSH-RSFNet achieves accuracy, precision, recall and F1-score of 95.51%, 95.15%, 95.30% and 95.22%, surpassing both baseline and recent CNN-based approaches, 3) compared to DSH-RSFNet, our proposed DSH-RSF-Lite approach obtains a significant decrease in model parameters, storage space, floating point operations and latency by more than 68.20%, 67.72%, 69.46% and 22.68%, while it has only 0.64%, 0.14%, 1.28% and 0.75% reduction in accuracy, precision, recall and F1-score, respectively. Extensive comparative experiments validate the effectiveness of our proposed DSH-RSF-Lite approach, highlighting its efficiency in augmenting chest X-ray images, improving the performance of pneumonia classification

关键词： Median filters

来源：评论

学校读者我要写书评

暂无评论

Dual Progressive Transformations for Weakly Supervised Semantic Segmentation

arXiv

引用

arXiv 2022年

作者： Huo, Dongjian Su, Yukun Wu, Qingyao School of Software and Engineering South China University of Technology China Key Laboratory of Big Data and Intelligent Robot Ministry of Education China Nanyang Technological University Singapore Pazhou Lab Guangzhou China

Weakly supervised semantic segmentation (WSSS), which aims to mine the object regions by merely using class-level labels, is a challenging task in computer vision. The current state-of-the-art CNN-based methods usually adopt Class-Activation-Maps (CAMs) to highlight the potential areas of the object, however, they may suffer from the part-activated issues. To this end, we try an early attempt to explore the global feature attention mechanism of vision transformer in WSSS task. However, since the transformer lacks the inductive bias as in CNN models, it can not boost the performance directly and may yield the over-activated problems. To tackle these drawbacks, we propose a Convolutional Neural Networks Refined Transformer (CRT) to mine a globally complete and locally accurate class activation maps in this paper. To validate the effectiveness of our proposed method, extensive experiments are conducted on PASCAL VOC 2012 and CUB-200-2011 datasets. Experimental evaluations show that our proposed CRT achieves the new state-of-the-art performance on both the weakly supervised semantic segmentation task the weakly supervised object localization task, which outperform others by a large margin. Code is available at https://***/huodongjian0603/crt Copyright © 2022, The Authors. All rights reserved.

关键词： Semantic Segmentation

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：