Kernel-based adversarial attacks and defenses on support vector classification

While malicious samples are widely found in many application fields of machine learning, suitable countermeasures have been investigated in the field of adversarial machine learning. Due to the importance and popularity of Support Vector Machines (SVMs), we first describe the evasion attack against SVM classification and then propose a defense strategy in this paper. The evasion attack utilizes the classification surface of SVM to iteratively find the minimal perturbations that mislead the nonlinear classifier. Specially, we propose what is called a vulnerability function to measure the vulnerability of the SVM classifiers. Utilizing this vulnerability function, we put forward an effective defense strategy based on the kernel optimization of SVMs with Gaussian kernel against the evasion attack. Our defense method is verified to be very effective on the benchmark datasets, and the SVM classifier becomes more robust after using our kernel optimization scheme.     

An intelligent and privacy-enhanced data sharing strategy for blockchain-empowered Internet of Things

With the development of the Internet of Things (IoT), the massive data sharing between IoT devices improves the Quality of Service (QoS) and user experience in various IoT applications. However, data sharing may cause serious privacy leakages to data providers. To address this problem, in this study, data sharing is realized through model sharing, based on which a secure data sharing mechanism, called BP2P-FL, is proposed using peer-to-peer federated learning with the privacy protection of data providers. In addition, by introducing the blockchain to the data sharing, every training process is recorded to ensure that data providers offer high-quality data. For further privacy protection, the differential privacy technology is used to disturb the global data sharing model. The experimental results show that BP2P-FL has high accuracy and feasibility in the data sharing of various IoT applications.     

A classification-based privacy-preserving decision-making for secure data sharing in Internet of Things assisted applications

The introduction of the Internet of Things (IoT) paradigm serves as pervasive resource access and sharing platform for different real-time applications. Decentralized resource availability, access, and allocation provide a better quality of user experience regardless of the application type and scenario. However, privacy remains an open issue in this ubiquitous sharing platform due to massive and replicated data availability. In this paper, privacy-preserving decision-making for the data-sharing scheme is introduced. This scheme is responsible for improving the security in data sharing without the impact of replicated resources on communicating users. In this scheme, classification learning is used for identifying replicas and accessing granted resources independently. Based on the trust score of the available resources, this classification is recurrently performed to improve the reliability of information sharing. The user-level decisions for information sharing and access are made using the classification of the resources at the time of availability. This proposed scheme is verified using the metrics access delay, success ratio, computation complexity, and sharing loss.     

CPFinder: Finding an unknown caller's profession from anonymized mobile phone data

Identifying an unfamiliar caller's profession is important to protect citizens' personal safety and property. Owing to the limited data protection of various popular online services in some countries, such as taxi hailing and ordering takeouts, many users presently encounter an increasing number of phone calls from strangers. The situation may be aggravated when criminals pretend to be such service delivery staff, threatening the user individuals as well as the society. In addition, numerous people experience excessive digital marketing and fraudulent phone calls because of personal information leakage. However, previous works on malicious call detection only focused on binary classification, which does not work for the identification of multiple professions. We observed that web service requests issued from users' mobile phones might exhibit their application preferences, spatial and temporal patterns, and other profession-related information. This offers researchers and engineers a hint to identify unfamiliar callers. In fact, some previous works already leveraged raw data from mobile phones (which includes sensitive information) for personality studies. However, accessing users' mobile phone raw data may violate the more and more strict private data protection policies and regulations (e.g., General Data Protection Regulation). We observe that appropriate statistical methods can offer an effective means to eliminate private information and preserve personal characteristics, thus enabling the identification of the types of mobile phone callers without privacy concerns. In this paper, we develop CPFinder —- a system that exploits privacy-preserving mobile data to automatically identify callers who are divided into four categories of users: taxi drivers, delivery and takeouts staffs, telemarketers and fraudsters, and normal users (other professions). Our evaluation of an anonymized dataset of 1,282 users over a period of 3 months in Shanghai City shows that the CPFinder can achieve accuracies of more than 75.0% and 92.4% for multiclass and binary classifications, respectively.     

A survey on deploying mobile deep learning applications: A systemic and technical perspective

With the rapid development of mobile devices and deep learning, mobile smart applications using deep learning technology have sprung up. It satisfies multiple needs of users, network operators and service providers, and rapidly becomes a main research focus. In recent years, deep learning has achieved tremendous success in image processing, natural language processing, language analysis and other research fields. Despite the task performance has been greatly improved, the resources required to run these models have increased significantly. This poses a major challenge for deploying such applications on resource-restricted mobile devices. Mobile intelligence needs faster mobile processors, more storage space, smaller but more accurate models, and even the assistance of other network nodes. To help the readers establish a global concept of the entire research direction concisely, we classify the latest works in this field into two categories, which are local optimization on mobile devices and distributed optimization based on the computational position of machine learning tasks. We also list a few typical scenarios to make readers realize the importance and indispensability of mobile deep learning applications. Finally, we conjecture what the future may hold for deploying deep learning applications on mobile devices research, which may help to stimulate new ideas.     

用于视网膜血管分割的半监督深度学习框架北大核心CSCD

针对目前视网膜血管分割任务中伪标签质量参差不齐,获得高质量的伪标签需要经过筛选的问题,本文提出了一种新的用于视网膜血管分割的半监督深度学习框架。该框架采用分而治之的思想来处理数据,针对有标签数据,采用传统的深度学习方法;针对无标签数据,采用Mean teacher模型,通过对比同一输入的不同形态输出,让模型学习无标签数据之间的共同特征,避免了采用伪标签技术带来的筛选过程。本文将U型网络(u-neural networks,U-Net)、Dense-Net和Ladder-Net 3个基准网络放入该框架,在DRIVE和CHASEDB1数据集上进行训练测试,均取得了较好的分割效果,表明本文框架具有提高网络区分不同阈值像素的能力。     

基于拓扑链路识别的光网络流量数据合成算法

基于深度学习的光网络流量诊断与预测等场景中,由于保密等原因,光链路的流量数据采集和存储工作受限。针对数据量少而无法支撑深度学习的问题,文章提出了一种基于拓扑链路识别的光网络流量数据合成算法,其核心思想是在生成对抗网络框架下,联合基于光网络拓扑的条件生成模型和基于光网络流量的数据合成模型,以自监督的方式合成指定光链路的流量数据。仿真结果表明,所提算法合成的光网络流量数据在自相关系数指标上与真实数据接近且使得基于全连接神经网络的流量预测模型准确率达到95%以上。     

基于策略学习的机票动态定价算法

机票动态定价旨在构建机票售价策略以最大化航班座位收益.现有机票定价算法都建立在提前预测各票价等级的需求量基础之上,会因票价等级需求量的预测偏差而降低模型性能.为此,提出基于策略学习的机票动态定价算法,其核心是不再预测各票价等级的需求量,而是将机票动态定价问题建模为离线强化学习问题.通过设计定价策略评估和策略更新的方式,从历史购票数据上学习具有最大期望收益的机票动态定价策略.同时设计了与现行定价策略和需求量预测方法的对比方法及评价指标.在两趟航班的多组定价结果表明:相比于现行机票销售策略,策略学习算法在座位收益上的提升率分别为30.94％和39.96％,且比基于需求量预测方法提升了6.04％和3.36％.     

基于深度学习的基金客服情绪安抚算法

构建基金智能客服情绪安抚算法,增加人性化设计以有效提升客户使用体验。由基于OCC模型的情绪粗粒度分类得到有负面情感的文本信息,对该文本信息进行相关处理,而后基于SVM模型进行情绪细粒度分类,得到文本的负面情绪类别并在囊括专业知识和情绪应答的基金领域知识库中匹配答案和情绪应答语句,将查询结果展示给客户。     

基于深度学习的ViSAR多运动目标检测

视频合成孔径雷达(ViSAR)在地面动目标检测和感兴趣区域(ROI)的动态监测方面具有巨大的潜力。对地面运动目标的检测与跟踪一直是ViSAR的研究热点。针对现有基于深度学习的ViSAR动目标检测方法存在的依赖预训练模型,模型迁移难等问题,本文提出了一种基于深度学习与多目标跟踪(MOT)算法的ViSAR动目标阴影检测方法。该方法首先设计了一种从零开始深度学习的网络模型,实现动目标阴影的单帧检测。为了提高检测性能的鲁棒性,采用了基于卡尔曼滤波和逐帧数据关联的多目标跟踪算法跟踪动目标。实测数据处理结果表明该方法具有良好的检测性能。