2020 August 08 Essay

读论文，记笔记

论文英语句子笔记

Article: Cooling-Shrinking Attack Blinding the Tracker with Imperceptible Noises

  1. <font color=#008000>Although several works have focused on attacking</font> image classifiers and object detectors,

<font color=#008000>an effective and efficient method for attacking</font> single object trackers of any target in a model-free way remains lacking.

  1. Research on adversarial attack is beneficial to understand deep neural networks and design robust models.

  2. The latter method is faster than the former method because only one-time forward propagation is needed for each attack after training.

  3. In the recent year, adversarial attack has become a popular topic and has extended from image classification to more challenging tasks, such as object detection [35, 31] and semantic segmentation [35].

  4. The latter type is generally faster than the former type because the operation for transforming an image to an adversarial one does not need to be repeated.

  5. AdvGAN [34] is the first work to generate adversarial examples with GAN, which can run efficiently during the inference phase.

  6. UEA achieves comparative performance but is much faster than DAG, because only one-time forward propagation is needed during the inference phase after training.(基于GAN的对抗样本生成技术优于其他方法的解释)

  7. Inheriting similar ideas （继承自该思想） 与 Inspired by 。。。一致

  8. 将自己方法时，首先用一句话指出我们的工作，或提出的方法；然后再指出提出该方法的目标；最后提出为了实现该目标，做了的工作。 In this work, we propose an adversarial perturbationgenerator for deceiving the SiamRPN++ tracker. The goal of our method is to make the target invisible to trackers,thereby leading to tracking drift. To accomplish this goal, we train the generator with a carefully designed and novel cooling-shrinking loss. Considering that SiamRPN-based trackers [16, 40, 15, 39, 30] locate the target in a local search region based on the template given in the initial frame, we design two versions of perturbation-generators to attack the search regions and the template respectively.

Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey

  1. Whereas deep learning performs a wide variety of Computer Vision tasks with remarkable accuracies

  2. Notice that the perturbations are hardly perceptible for human vision system, however their effects on the deep learning models are catastrophic.

  3. The remaining article also follows the same definitions of the terms. 表述某个符号或者定义在下文不发生变化

  4. laboratory setting, in the digital world.

Matching Networks for One Shot Learning

  1. 规则的单词：criterion, regime, strategy
  2. more precisely, more specifically
  3. improve …. on dataset xx% to xx% and from xx% to xx% dataset compared to competing approaches.
  4. The novelty of our work is twoflod:
  5. we draw inspiration from models such as sequence to sequence with attention.
  6. Unless otherwise specified, training is always on, and test in .. 除非另有说明，
  7. as reported in their paper 引用他人文章数据用语
  8. An obvious drawback of our model is the fact that, 表述方法的缺陷

One-shot Learning with Memory-Augmented Neural Networks

  1. adequately incorporate 充分纳入
  2. catastrophic interference 灾难性推断
  3. obviate the downside of 消除…的不利因素
  4. The network is tasked to…/The goal of the network is …网络的任务是/网络的目标是
  5. akin to 类似于
  6. As such 因此
  7. prior to… 先于
  8. in accordance with 依据（后面接公式等）

New High-speed Frame Detection Method: Side Trace Algorithm (STA) for i-appli on Cellular Phones to Detect Watermarks

  1. the response delay caused by the transmission time presented a problem for the user interface. 由传输时间引起的相应延迟给用户交互引出了一个难题。 句式：名词（现象）+ 原因状语 + 谓语 + 宾语 （个人己见）
  2. The drawback to this method is that 这个方法的缺点是
  3. the former , 二者中的前者
  4. A feature point search must be able to withstand the effects of this noise 特征点搜索必须能够经受得了这种噪声。
  5. quadrilateral 四边形
  6. bisect 二等分
  7. traversal 遍历

Screen-Shoting Resilient Watermarking

  1. traversal 遍历
  2. The reason can be analyzed as follows. 原因解释如下
  3. intensity 强度
  4. As explained previously , 正如之前的解释
  5. The second requirement is motivated by the observation that …. 第二个要求是由以下观察得出：
  6. Similar conclusion can be drawn as that for Dataset8. 可以得出与数据集8相似的实验结论
  7. It is observed that, 可以观察到

The Unreasonable Effectiveness of Deep Features as a Perceptial Metric

  1. The ability to compare data items is perhaps the most fundamental operation underlying all of computing.

    句意：比较数据项的能力也许能所有计算的基础的最基本操作

  2. the goal is for the compressed image to be indistinguishable from the original by a human observer， irrespective of the fact that their pixel representations might be very different. 句意：压缩图像的目标是尽可能让一个人类观察者感知不到与原始图像的差别，尽管实际上它们像素表示上有一些差别。

  3. A well-know example is that blurring causes large perceptual but small $\ell_2$ change. 句意：一个显而易见的例子是感知上造成较大模糊的情况下，二范数损失的该变量却很小。

  4. Our results are consistent with the hypothesis that perceptual similarity is not a special function all of its own, but rather a consequence of visual representations tuned to be predictive about important structure in the world. 句意：我们的结果与假设一致：感知相似性并不是其本身的特殊功能，而是视觉表示的结果，这些视觉表示被调整为可预测现实世界的重要结果。

  5. We create a set of “traditional” distortions consisting of common operations performed on the input pathches, listed in Table 1. In general, we use photometric distortions, random noise, blurring, spatial shifts and corruptions, and compression artifacts. 句意：我们创建了一个”传统“失真集合，包括在输入patches上常用的一些操作。一般而言，我们使用。。。。
  6. In addition, we collect judgements on a large number of 64 x 64 patches rather than a small number of images. There are three reasons for this. First,… 句意：此外，我们在64x64尺寸的patches集合中收集判断，而不是小数据集。原因是。。。
  7. facto standard 事实标志

Text-image Watermarking based on Integer Wavelet Transform(IWT) and Discrete Cosine Transform(DCT)

  1. Integer Wavelet Transform(IWT) is much faster beacuse it deals only with integers. The image can be reconstructed without any loss using IWT and can be stored without rounding off the error 1 句意：IWT更快，因为它仅处理整数。使用IWT重建图像而不会造成任何损失，并且可以存储图像出现错误？
  2. In this section, watermarking methods which performed on text-image will be discussed and compared. Text-image watermarking methods in the spatial domain are presented first, then watermarking methods which are performed in the transform domain. 句意：在本节，将会讨论数字水印用于文本图片的方法。首先介绍空域方法，然后介绍变换域方法。

Image Demoireing with Learnable Bandpass Filters

  1. Through an ablation study, we demonstrate the effectiveness of the different components of Net 句意：通过对比实研究，我们证明了网络的不同组件的有效性。
  2. **In this section, we make a breif introduction of **several CNN-based methods in related tasks, where deep learning has made significant impact. 句意：在本节，我们简短地介绍了基于CNN方法在相关任务中的工作，这些工作中深度学习方法取得了重大的进展。
  3. L1 loss is more effective than L2 loss for image restoration tasks. 句意：在图像重建任务中，L1损失的比L2损失更加有效
  4. Please note: the validation dataset is completely independent and not used in training. 句意：请注意：验证集完全独立并且没有在训练中使用 这个可以视情况使用，比如说实验结果都在什么数据集上的
  5. To verify the effectiveness of each component in our model, we conduct extensive ablation study, including evaluation of MTRB vs. GTMB and LTMB, learnable bandpass filter, and loss function. 这段是指出消融实验都包括哪些内容，可以在消融实验标题下的第一段使用。
    • The result is mainly caused by two reasons. First, …. ;(用.)Second,… 句意：解释某项结果的原因的句式。
    • There are two reasons for this observation. First, …. ;(用.)Second,… 句意：这个观察的原因有两个。
  7. The CNN has certain denoising and local smoothing capabilities. 句意：CNN有一定的去噪和局部平滑能力。
  8. We denote the networks constructed without LP and DDT as MBCNN-nLP and MBCNN-nDDT, respectively. 用法：在对网络进行修改的对比实验设置时，可以使用该句法。 句意：我们将不带LP或DDT的网络结构分别称为。。。。
  9. We tested the performance of these three models on validation set of LCDMoire. 句意：我们在验证机上测试了三个模型的性能。

##

  1. whether the model assigns the perturbed input to some class different from the class label of original input.
  2. Target misclassification for which the perturbed input has to be classified in a given target class.

VisCode embedding information in visualization images using encoder-decoder network

  1. disseminating 传播
  2. salient feature 显著性特征
  3. The resolution of the output coded image is the same as that of the input carrier image. 句意：输出编码图的分辨率与输入载体图一致
  4. We refer the reader to the ISO/IEC 18004 specification for more technical details of QR codes. 句意： 在推荐让读者了解某件事的时候可以使用

Drug repurposing: progress, challenges and recommendations

  1. drug repurposing 药物再利用。 drug repositioning 药物重定位
  2. attrition 损耗
  3. substantial 充实的，坚实，实惠，殷实，物质的
  4. escalating 逐步升级，升高
  5. in generic terms, 一般而言
  6. underlying/latent features 潜在特征

Adversarial Learning for Invertible Steganography

  1. pristine 原始的
  2. to name but a few 这仅仅是列举的一小部分
  3. Experimental result validated the effectiveness of the proposed method and showed a significant performance boost. 实验结果验证了提出方法的有效性和显示了巨大的性能提升。
  4. revisits 回顾
  5. UNet 模型的介绍：It is characterised by it U-shape, which explains the origins of the name, and is comprised of a pair of encoder and decoder with skip connections between mirrored layers. The samples flow from top to bottom through a series of convolutions and downsampling, and then go back through a succession of convolutions and upscaling to full resolution. The skip connections allow multi-resolution feature maps from previous layers to be retained and concatenated with feature maps at later layers.
  6. l1 norm may be preferable to l2 norm for the case of high dimensional data 2. l1 norm is capable of capture low-frequency structure, if the model solely rests on the l1 norm, the outcome would trend to be blurry since this loss function is minimised by averaging all plausible outputs, thus incentivising a blur when uncertainty encountered in complex areas.
  7. 本文使用GAN完成可逆隐写的任务。具体地说，作者将图像转化成8个位图，根据正则奇数映射（Regular-singular, RS）的原理，将7个无失真的位平面输入到生成器里面，输出一个目标位平面（利用未修改的位平面来预测或合成修改的位平面）。该目标位平面被用于可逆隐写时嵌入和提取的参考位图。最后对生成器的合成图进行了后处理（量化成0-1数据）。疑问在于：1、神经网络无法做到完全不失真，这是否符合可逆隐写的规则？2、如果生成器只是用来生成目标参考位图，那么是如何实现隐写和信息提取的？在文中提到一句，“Blocks for carrying message bits was fixed to the size of 2 x 2 pixels”。 所以最终的信息位是否是合成图按2x2大小像素块计算得到的信息？所以最终的逻辑是：输入7个不变的位平面得到可改变的目标位平面，得到目标位平面后呢？与原来7个不变的位平面叠加重建图片？如何实现可逆呢？

Functional adversarial attack (NeurIPS 2019)

  1. cause the algorithm to report(替换make) an erroneous（替换wrong） outpout, e.g. the incorrect（替换wrong） label for classifier.
  2. a change in the environment that is not noticeable to a human may cause unexpected, unwanted, or dangerous behavior.
  3. enforces/guide 逼迫/指导

Unsupervised Domain Adaptation using Feature-Whitening and Consensus Loss (CVPR 2019)

领域自适应存在的两个问题：a) 源域和目标域不匹配； b) 无标注目标数据假设，超参数在UDA中很难调优。

UDA方法可以分成以下几类：

熵最小化是一种在半监督领域种常见的策略。它利用未标注样本的高置信度作为伪标注。后来在很多UDA中都能看到。

其他UDA中的范式，称为一致性逼迫范式，然后在相同目标输入的两个扰动版本的预测之间强加一些一致性。通过设计一些损失函数来实现这样的一致性。

方法

常见的技术是在网络中使用基于BN的层来减少领域偏移，比如通过特征标准化将源域和目标域分布投影到一个参考域中。作者使用whitening实现特征标准化。使用Min-Entropy Consensus(MEC)损失，网络的输入要由两个编程三个。

特征whitening完全去批样本的相关性，因此潜在地提升损失函数的landsacpe的平滑度。在无标签数据集上时，损失函数的平滑重要性高于类似熵的损失。

Accessibility 可访问性

On Adaptive Attacks to Adversarial Example Defenses (CVPR 2020)（对抗样本防御的自适应攻击）

词汇
  1. de facto 实际上
  2. bootstrap 自展

Reference

[1] R. Harshitha, S.S., Vidya, Robust and High Limit Watermarking using DWT-IWT. International Journal of Advanced Scientific Research and Management, 2 (4), pp.18-21, 2017.

[2]C. C. Aggarwal, A. Hinneburg, and D. A. Keim, “On the surprising behavior of distance metrics in high dimensional space,” in Proceedings of International Conference on Database Theory (ICDT), London, UK, 2001, pp. 420–434.

All content is licensed under CC BY-NC-SA

Built with Jekyll and 3-Jekyll theme • Hosted on Github

Table of Contents