poisoning_certified_defenses:基于随机平滑的数据中毒防御有多强健? 出现在CVPR 2021中

上传者: 42173218 | 上传时间: 2023-01-06 00:09:13 | 文件大小: 57KB | 文件类型: ZIP
基于随机平滑的数据中毒防御有多强健? 抽象的 可证明可靠的分类器的预测在一个点附近保持不变,从而使它们在保证测试时间的情况下具有弹性。 在这项工作中,我们提出了对健壮的机器学习模型的前所未有的威胁,突显了训练数据质量在实现高认证健壮性方面的重要性。 具体而言,我们提出了一种基于双层优化的新型数据中毒攻击,该攻击会降低可证明的鲁棒分类器的鲁棒性保证。 与其他数据中毒攻击会降低一小组目标点上的中毒模型的准确性不同,我们的攻击会减少数据集中整个目标类的平均认证半径。 此外,即使受害者使用最新的健壮训练方法(例如, 和从头开始训练模型,我们的攻击也是有效的。 为了使攻击更难检测,我们使用带有明显较小失真的干净标签中毒点。 通过中毒MNIST和CIFAR10数据集并使用前面提到的鲁棒训练方法训练深度神经网络,并使用随机平滑验证其鲁棒性,来评估所提出方法的有效性。 对于使用这些强大的训练方法训练的模型

文件下载

资源详情

[{"title":"( 25 个子文件 57KB ) poisoning_certified_defenses:基于随机平滑的数据中毒防御有多强健? 出现在CVPR 2021中","children":[{"title":"poisoning_certified_defenses-main","children":[{"title":"models","children":[{"title":"mnist_keras_models.py <span style='color:#111;'> 822B </span>","children":null,"spread":false},{"title":"cifar10_keras_models.py <span style='color:#111;'> 8.15KB </span>","children":null,"spread":false}],"spread":true},{"title":"attack_generation","children":[{"title":"bilevel_optimizer","children":[{"title":"bilevel_approxgrad.py <span style='color:#111;'> 2.78KB </span>","children":null,"spread":false}],"spread":true},{"title":"macer","children":[{"title":"CIFAR10","children":[{"title":"poison_radius_macer.py <span style='color:#111;'> 6.17KB </span>","children":null,"spread":false},{"title":"bilevel_poison_radius_macer.py <span style='color:#111;'> 11.95KB </span>","children":null,"spread":false}],"spread":true},{"title":"MNIST","children":[{"title":"poison_certified_radius_macer.py <span style='color:#111;'> 5.96KB </span>","children":null,"spread":false},{"title":"bilevel_poison_radius_macer.py <span style='color:#111;'> 11.60KB </span>","children":null,"spread":false}],"spread":true}],"spread":true},{"title":"gaussian_data_augmentation","children":[{"title":"CIFAR10","children":[{"title":"bilevel_poison_radius.py <span style='color:#111;'> 7.32KB </span>","children":null,"spread":false},{"title":"poison_radius_gaussianaug.py <span style='color:#111;'> 5.19KB </span>","children":null,"spread":false}],"spread":true},{"title":"MNIST","children":[{"title":"poison_radius_gaussianaug.py <span style='color:#111;'> 5.40KB </span>","children":null,"spread":false},{"title":"bilevel_poison_radius_gaussianaug.py <span style='color:#111;'> 7.66KB </span>","children":null,"spread":false}],"spread":true}],"spread":true},{"title":"smoothadv","children":[{"title":"CIFAR10","children":[{"title":"poison_radius_smoothadv.py <span style='color:#111;'> 11.62KB </span>","children":null,"spread":false},{"title":"bilevel_poison_radius_smoothadv.py <span style='color:#111;'> 6.93KB </span>","children":null,"spread":false}],"spread":true},{"title":"MNIST","children":[{"title":"bilevel_poisoning_smoothadv_pgd.py <span style='color:#111;'> 6.93KB </span>","children":null,"spread":false},{"title":"poison_certified_radius_smoothadv_pgd.py <span style='color:#111;'> 11.49KB </span>","children":null,"spread":false}],"spread":true}],"spread":true}],"spread":true},{"title":"evaluation","children":[{"title":"CIFAR10","children":[{"title":"smoothadv_training.py <span style='color:#111;'> 7.33KB </span>","children":null,"spread":false},{"title":"gaussian_augmented_training.py <span style='color:#111;'> 4.40KB </span>","children":null,"spread":false},{"title":"macer_training.py <span style='color:#111;'> 6.20KB </span>","children":null,"spread":false}],"spread":true},{"title":"randomized_smoothing_certification","children":[{"title":"certify_mnist.py <span style='color:#111;'> 5.71KB </span>","children":null,"spread":false},{"title":"certify_cifar10.py <span style='color:#111;'> 5.86KB </span>","children":null,"spread":false},{"title":"core.py <span style='color:#111;'> 5.16KB </span>","children":null,"spread":false}],"spread":true},{"title":"MNIST","children":[{"title":"smoothadv_training.py <span style='color:#111;'> 7.03KB </span>","children":null,"spread":false},{"title":"gaussian_augmented_training.py <span style='color:#111;'> 3.88KB </span>","children":null,"spread":false},{"title":"macer_training.py <span style='color:#111;'> 5.86KB </span>","children":null,"spread":false}],"spread":true}],"spread":true},{"title":"README.md <span style='color:#111;'> 6.18KB </span>","children":null,"spread":false}],"spread":true}],"spread":true}]

评论信息

免责申明

【只为小站】的资源来自网友分享,仅供学习研究,请务必在下载后24小时内给予删除,不得用于其他任何用途,否则后果自负。基于互联网的特殊性,【只为小站】 无法对用户传输的作品、信息、内容的权属或合法性、合规性、真实性、科学性、完整权、有效性等进行实质审查;无论 【只为小站】 经营者是否已进行审查,用户均应自行承担因其传输的作品、信息、内容而可能或已经产生的侵权或权属纠纷等法律责任。
本站所有资源不代表本站的观点或立场,基于网友分享,根据中国法律《信息网络传播权保护条例》第二十二条之规定,若资源存在侵权或相关问题请联系本站客服人员,zhiweidada#qq.com,请把#换成@,本站将给予最大的支持与配合,做到及时反馈和处理。关于更多版权及免责申明参见 版权及免责申明