2021年09月_Billy1900

09月 07月 06月 04月

原创 DeepInspect: A Black-box Trojan Detection and Mitigation Framework for Deep Neural Networks笔记

Challenges:C1.后门的隐蔽性使得他们很难通过功能测试functional testing来识别（这种测试通常会使用测试准确率作为检测标准）C2.在后门检测期间，只能得到被查询模型的有限信息。在真实世界中一个干净的训练数据集或者一个好的推理模型可能不容易得到。训练集包含用户的个人信息，所以一般不会分布在预训练的模型中。C3.防御者不知道攻击者指定的target。在我们的情况下，攻击者是恶意模型的提供者而防守者是终端用户。攻击者目标的未知性使得NT的检测更加复杂，因为对于有大量输出clas.

2021-09-03 16:23:25 539

原创 Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks 笔记

Code: https://github.com/kangliucn/Fine-pruning-defenseContext: 因为在我们训练模型时，由于我们所需要的数据集太大，对于计算机的配置要求太高，时间消耗太大，所以我们一般会将模型的训练过程外包给第三方，而会存在着有着恶意想法的第三方，会在给我们训练的模型上安上后门，影响我们模型的判断，这时候就需要我们运用不同的方法来‘消毒’。作者自己给自己设计了新型的后门攻击并且用了自己设计的fine-pruning（a combination of prun

2021-09-03 16:20:47 584

原创 STRIP: A Defence Against Trojan Attacks on Deep Neural Networks 笔记

Code: https://github.com/garrisongys/STRIP?utm_source=catalyzex.com本文提出的 detect 方法如下图所示：说明：（图中 input x 是待检测的样本，这里示例的是一张加入了 trigger（右下角的小方块）的样本，作者将验证集中的干净图片依次作为水印添加到 x 上，得到多张添加了水印的 perturbed inputs（比如作者用了 100 张干净图片做水印，那么依次添加后就得到 100 perturbed inputs

2021-09-03 16:19:28 955

原创 Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks笔记

Code: https://github.com/mzweilin/EvadeML-ZooFeature squeezing: reducing the color bit depth of each pixel and spatial smoothing.Framework:Adversarial examples attacksLpnormL_p normLpnorm attackFGSMBIMDeepFoolJSMACarlini/Wagner attacksD

2021-09-03 16:17:58 1091

原创 Backdoor Attack with Sample-Specific Triggers (2021) 笔记

Backdoor attack injects some attacker-specified patterns in the poisoned image and replace the corresponding label with a pre-defined target label.Existing backdoor attacks are easily defensed because the backdoor triggers are sample-agnosticwe ge..

2021-09-03 16:15:16 522

原创 TrojanNet: Embedding Hidden Trojan Horse Models in Neural Networks (2020)笔记

Code: https://github.com/Billy1900/TrojanNet (self produced)We prove theoretically that the Trojan network’s detection is computationally infeasible and demonstrate empirically that the transport network does not compromise its disguise.Our method ut

2021-09-03 16:13:52 813

原创 BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain (2017)笔记

Code: https://github.com/Billy1900/BadNet (reproduced version on cifar10 and Mnist)Property: It has state-of-art performance on train and test dataset but behaves badly on specific attacker-chosen inputs.Context: outsourced task and transfer learningCa

2021-09-03 16:02:26 703

空空如也

TA创建的收藏夹 TA关注的收藏夹

TA关注的人

Billy1900的博客