ID 原文 译文
56168 该任务的核心挑战在于网络爬取的图像中存在一定量的类别噪声,从而影响自主学习的可靠性. A major challenge in this task is the existence of label noise in web images.
56169 为了解决类别噪声问题,本文设计了一种新颖的噪声擦除模型.该模型通过每次从小批次样本的置信注意力区域中以跨样本的方式学习语义信息来擦除训练图像中与搜索关键词无关的区域. To dealwith the label noise, we design a noise erasing network that is able to learn cross-image knowledge from credibleattention regions in images in a mini-batch and then erase those regions unrelated to the search keywords fromthe web images. With this network, our system can automatically generate high-quality ‘proxy ground truth’, fortraining semantic segmentation models.
56170 基于该模型,本文同时提出了一种能够用于训练语义分割模型的高质量伪标注生成方法. Extensive experiments on the popular benchmark, i. e.
56171 在国际主流的公开数据集(PASCAL VOC2012)上的大量实验表明,基于该方法的语义分割模型在利用网络监督与弱监督的条件下均取得了良好结果 (mIoU=62. 0%以及66. 1%). , PASCAL VOC2012, show surprisingly good results in both our task (mIoU = 62. 0%) and the weakly-supervised setting (mIoU= 66. 1%).
56172 移动设备的普及带来了移动应用程序市场的蓬勃发展,各类服务提供商通过移动应用程序的权限大量收集用户数据,而数据收集过程往往不为用户所知,因此给用户带来极大的隐私风险. With the prevalence of mobile devices and mobile applications (Apps), service providers have becomeincreasingly enthusiastic in collecting user data, which would cause huge privacy risk due to the invisibility ofdata collection.
56173 对移动应用程序进行隐私风险评估,不仅有助于规范第三方移动应用市场,而且可帮助用户规避潜在的隐私风险,而如何评估移动应用程序可能带来的最大隐私风险则是当前面临的重大挑战. How to evaluate the maximum privacy risks of mobile Apps is a key challenge, which notonly contributes to the regulation of App market, but also helps users to avoid potential privacy leakage.
56174 本文通过研究移动应用程序最大化的数据泄露场景,基于权限请求特征和权限分析原则构建隐私风险最大值量化模型. Byinvestigating the maximum data leakage of an App, this paper proposes a privacy risk quantification model basedon the requested permissions and the principles of App permission analysis.
56175 该模型基于权限敏感度、权限类别异常度、权限使用率和权限调用者数量4个参数,对移动应用程序的潜在隐私风险进行评估. The proposed model introduces fourimportant parameters, namely, permission sensitivity, anomaly degree of permission list, utilization rate of anApp, and number of permission callers, to evaluate the potential privacy risk of an App.
56176 在隐私风险量化和恶意应用检测中,对比当前同类型方法,该模型在真实数据集上效果均较优,说明模型的有效性. We conduct experimentsof privacy risk evaluation and malicious App detection over real datasets, and the results show that our proposedmodel achieves better performance against state-of-the-art solutions, which demonstrates the effectiveness of thismodel.
56177 实验结果进一步表明,该模型可用于改善现有第三方移动应用市场的隐私风险预警机制,进而保护移动用户的隐私. Further, analytical results also indicate that this privacy risk quantification model can serve as an effectiveprivacy risk warning mechanism for user privacy preservation.