【论文阅读笔记】CRFL: Certifiably Robust Federated Learning against Backdoor Attacks

个人阅读笔记,如有错误欢迎指出!

会议:PMLR 2021[2106.08283] CRFL: Certifiably Robust Federated Learning against Backdoor Attacks (arxiv.org)

问题:

        现有的防御算法缺乏健壮性

创新:

        证明了所提出框架得稳定性

        通过马尔可夫核分析聚合模型的训练过程,提出用于模型推理得参数平滑

方法:

        攻击方:single-shot,同时攻击

        1、训练阶段:裁剪范数 添加扰动

                服务器对收到的客户端范数通过阈值\rho_t​进行剪裁Clip_{\rho_t}\gets w_t/max(1,\frac{||w_t||}{\rho_t})

                对聚合后的全局模型添加高斯噪声\epsilon_t \sim \mathcal{N}(0,\sigma^2_I)

                则最终融合后的参数为\widetilde{w}_t \gets Clip_{\rho_t}(w_t)+\epsilon_t

                在最后一轮中只剪裁全局模型参数

                算法流程

        2、测试阶段:参数平滑

                构建平滑分类器h:(\mathcal{W},\mathcal{X}\rightarrow \mathcal{Y}),并基于此分类器预测

                对原始全局模型的预测结果进行投票,获取概率最大的class(得票最多的类)

                测试期间对裁剪后的全局模型\mathcal{M}次添加高斯噪声\mu(w)=\mathcal{N}(w,\sigma^2_TI),用来估计\mathcal{M}个蒙特卡洛样本(近似类别概率\hat{p_c}​)

                GetCounts

                用测试样本x_{test}​的每组噪声模型参数w^k_T,k\in\mathcal{M}运行分类器,并返回计数向量

                选择最大的两类\hat{c}_a\hat{c}_b​并计算其相应的\hat{p_A}\hat{p_B}

                CalculateBound经验校准估计

                调整经验估计,约束平滑模型返回错误标签的概率,误差容忍度为\alpha

                使用Hoeffding不等式计算\hat{p_A}​的下界和\hat{p_B}的上界

                算法流程

        与中心化设置(RAB)中可证明鲁棒性比较:RAB使用M个噪声扰动的数据来训练M个模型,属于输入数据扰动。CRFL只训练了一个全局模型,并且最终生成M个噪声扰动的模型副本,输入模型参数扰动。

实验:

⼈⼯智能会议等级列表 版权声明:本⽂为博主原创⽂章,遵循版权协议,转载请附上原⽂出处链接和本声明。 本⽂链接: 中国计算机学会推荐国际学术会议 (⼈⼯智能与模式识别) ⼀、A类 序号 会议简称 会议全称 出版社 ⽹址 1 AAAI AAAI Conference on Artificial Intelligence AAAI 2 CVPR IEEE Conference on Computer Vision and Pattern Recognition IEEE 3 ICCV International Conference on Computer Vision IEEE 4 ICML International Conference on Machine Learning ACM 5 IJCAI International Joint Conference on Artificial Intelligence Morgan Kaufmann ⼆、B类 序号 会议简称 会议全称 出版社 ⽹址 1 COLT Annual Conference on Computational Learning Theory Springer 2 NIPS Annual Conference on Neural Information Processing Systems MIT Press 3 ACL Annual Meeting of the Association for Computational Linguistics ACL 4 EMNLP Conference on Empirical Methods in Natural Language Processing ACL 5 ECAI European Conference on Artificial Intelligence IOS Press 6 ECCV European Conference on Computer Vision Springer 7 ICRA IEEE International Conference on Robotics and Automation IEEE and Automation 8 ICAPS International Conference on Automated Planning and Scheduling AAAI 9 ICCBR International Conference on Case-Based Reasoning Springer 10 COLING International Conference on Computational Linguistics ACM 11 KR International Conference on Principles of Knowledge Representation and Reasoning Morgan Kaufmann 12 UAI International Conference on Uncertainty in Artificial Intelligence AUAI 13 AAMAS International Joint Conference on Autonomous Agents and Multi-agent Systems Springer 三、C类 序号 会议简称 会议全称 出版社 ⽹址 1 ACCV Asian Conference on Computer Vision Springer 2 CoNLL Conference on Natural Language Learning CoNLL 3 GECCO Genetic and Evolutionary Computation Conference ACM 4 ICTAI IEEE International Conference on Tools with Artificial Intelligence IEEE 5 ALT International Conference on Algorithmic Learning Theory Springer 6 ICANN International Conference on Artificial Neural Networks Springer 7 FGR International Conference on Automatic Face and Gesture Recognition IEEE 8 ICDAR International Conference on Document Analysis and Recognition IEE
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值