Fate1.6 支持的机器学习算法

蔚落 2022-09-10 15:14 9阅读 0赞

# 联邦机器学习 #

Federatedml模块包括许多常见机器学习算法联邦化实现。所有模块均采用去耦的模块化方法开发，以增强模块的可扩展性。具体来说，我们提供：

1.  联邦统计: 包括隐私交集计算，并集计算，皮尔逊系数, PSI等
2.  联邦特征工程：包括联邦采样，联邦特征分箱，联邦特征选择等。
3.  联邦机器学习算法：包括横向和纵向的联邦LR, GBDT， DNN，迁移学习等
4.  模型评估：提供对二分类，多分类，回归评估，聚类评估，联邦和单边对比评估
5.  安全协议：提供了多种安全协议，以进行更安全的多方交互计算。

[![联邦学习结构][f7ff42452bc71e8f67317e166ba1f6ba.png]][f7ff42452bc71e8f67317e166ba1f6ba.png 1]

图1

## 算法清单 ##

<table> 
 <thead> 
  <tr> 
   <th>算法</th> 
   <th>模块名</th> 
   <th>描述</th> 
   <th>数据输入</th> 
   <th>数据输出</th> 
   <th>模型输入</th> 
   <th>模型输出</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>Reader</td> 
   <td>Reader</td> 
   <td>当输入数据的存储引擎当前计算引擎不支持时，会自动转存到FATE集群适配计算引擎的组件输出存储引擎；当输入数据的存储格式非FATE支持存储格式时，会自动转换格式，并存储到FATE集群的组件输出存储引擎</td> 
   <td>用户原始存储数据</td> 
   <td>转换后原始数据</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/util/README.html" rel="nofollow">DataIO</a></td> 
   <td>DataIO</td> 
   <td>该组件将原始数据转换为Instance对象(FATE-v1.7后会逐步弃用，使用DataTransform)。</td> 
   <td>Table，值为原始数据</td> 
   <td>转换后的数据表，值为在 : federatedml/feature/instance.py 中定义的Data Instance的实例</td> 
   <td></td> 
   <td>DataIO模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/README_zh.html#id6" rel="nofollow"><code>DataTransform</code>_</a></td> 
   <td>DataTransform</td> 
   <td>该组件将原始数据转换为Instance对象。</td> 
   <td>Table，值为原始数据</td> 
   <td>转换后的数据表，值为在 : federatedml/feature/instance.py 中定义的Data Instance的实例</td> 
   <td></td> 
   <td>DataTransform模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/intersect/README.html" rel="nofollow">Intersect</a></td> 
   <td>Intersection</td> 
   <td>计算两方的相交数据集，而不会泄漏任何差异数据集的信息。主要用于纵向任务。</td> 
   <td>Table</td> 
   <td>两方Table中相交的部分</td> 
   <td></td> 
   <td>Intersect模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Federated Sampling</a></td> 
   <td>FederatedSample</td> 
   <td>对数据进行联邦采样，使得数据分布在各方之间变得平衡。这一模块同时支持单机和集群版本。</td> 
   <td>Table</td> 
   <td>采样后的数据，同时支持随机采样和分层采样</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Feature Scale</a></td> 
   <td>FeatureScale</td> 
   <td>特征归一化和标准化。</td> 
   <td>Table，其值为instance</td> 
   <td>转换后的Table</td> 
   <td>变换系数，例如最小值/最大值，平均值/标准差</td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Hetero Feature Binning</a></td> 
   <td>Hetero Feature Binning</td> 
   <td>使用分箱的输入数据，计算每个列的iv和woe，并根据合并后的信息转换数据。</td> 
   <td>Table，在guest中有标签y，在host中没有标签y</td> 
   <td>转换后的Table</td> 
   <td></td> 
   <td>每列的iv/woe，分裂点，事件计数，非事件计数等</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Homo Feature Binning</a></td> 
   <td>Homo Feature Binning</td> 
   <td>计算横向场景的等频分箱</td> 
   <td>Table</td> 
   <td>转换后的Table</td> 
   <td></td> 
   <td>每列的分裂点</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">OneHot Encoder</a></td> 
   <td>OneHotEncoder</td> 
   <td>将一列转换为One-Hot格式。</td> 
   <td>Table, 值为Instance</td> 
   <td>转换了带有新列名的Table</td> 
   <td></td> 
   <td>原始列名和特征值到新列名的映射</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Hetero Feature Selection</a></td> 
   <td>HeteroFeatureSelection</td> 
   <td>提供多种类型的filter。每个filter都可以根据用户配置选择列。</td> 
   <td>Table, 值为Instance</td> 
   <td>转换的Table具有新的header和已过滤的数据实例</td> 
   <td>模型输入如果使用iv filters，则需要hetero_binning模型</td> 
   <td>每列是否留下</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/union/README.html" rel="nofollow">Union</a></td> 
   <td>Union</td> 
   <td>将多个数据表合并成一个。</td> 
   <td>Tables</td> 
   <td>多个Tables合并后的Table</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/linear_model/logistic_regression/README.html" rel="nofollow">Hetero-LR</a></td> 
   <td>HeteroLR</td> 
   <td>通过多方构建纵向逻辑回归模块。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>Logistic回归模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/local_baseline/README.html" rel="nofollow">Local Baseline</a></td> 
   <td>LocalBaseline</td> 
   <td>使用本地数据运行sklearn Logistic回归模型。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/linear_model/linear_regression/README.html" rel="nofollow">Hetero-LinR</a></td> 
   <td>HeteroLinR</td> 
   <td>通过多方建立纵向线性回归模块。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>线性回归模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/linear_model/poisson_regression/README.html" rel="nofollow">Hetero-Poisson</a></td> 
   <td>HeteroPoisson</td> 
   <td>通过多方构建纵向泊松回归模块。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>泊松回归模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/linear_model/logistic_regression/README.html" rel="nofollow">Homo-LR</a></td> 
   <td>HomoLR</td> 
   <td>通过多方构建横向逻辑回归模块。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>Logistic回归模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/nn/homo_nn/README.html" rel="nofollow">Homo-NN</a></td> 
   <td>HomoNN</td> 
   <td>通过多方构建横向神经网络模块。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>神经网络模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/ensemble/README.html" rel="nofollow">Hetero Secure Boosting</a></td> 
   <td>HeteroSecureBoost</td> 
   <td>通过多方构建纵向Secure Boost模块。</td> 
   <td>Table，值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>SecureBoost模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/ensemble/README.html" rel="nofollow">Hetero Fast Secure Boosting</a></td> 
   <td>HeteroFastSecureBoost</td> 
   <td>使用分层/混合模式快速构建树模型</td> 
   <td>Table，值为Instance</td> 
   <td>Table，值为Instance</td> 
   <td></td> 
   <td>FastSecureBoost模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/sbt_feature_transformer" rel="nofollow">Hetero Secure Boost Feature Transformer</a></td> 
   <td>SBT Feature Transformer</td> 
   <td>利用SBT叶子为特征编码</td> 
   <td>Table，值为Instance</td> 
   <td>Table，值为Instance</td> 
   <td></td> 
   <td>SBT Transformer模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/evaluation/README.html" rel="nofollow">Evaluation</a></td> 
   <td>Evaluation</td> 
   <td>为用户输出模型评估指标。</td> 
   <td>Table(s), 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/correlation/README.html" rel="nofollow">Hetero Pearson</a></td> 
   <td>HeteroPearson</td> 
   <td>计算来自不同方的特征的Pearson相关系数。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/nn/hetero_nn/README.html" rel="nofollow">Hetero-NN</a></td> 
   <td>HeteroNN</td> 
   <td>构建纵向神经网络模块。</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>纵向神经网络模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/ensemble/README.html" rel="nofollow">Homo Secure Boosting</a></td> 
   <td>HomoSecureBoost</td> 
   <td>通过多方构建横向Secure Boost模块</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>SecureBoost模型，由模型本身和模型参数组成</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Homo OneHot Encoder</a></td> 
   <td>横向 OneHotEncoder</td> 
   <td>将一列转换为One-Hot格式。</td> 
   <td>Table, 值为Instance</td> 
   <td>转换了带有新列名的Table</td> 
   <td></td> 
   <td>原始列名和特征值到新列名的映射</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/model_selection/data_split/README.html" rel="nofollow">Data Split</a></td> 
   <td>数据切分</td> 
   <td>将输入数据集按用户自定义比例或样本量切分为3份子数据集</td> 
   <td>Table, 值为Instance</td> 
   <td>3 Tables</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/feature/README.html" rel="nofollow">Column Expand</a></td> 
   <td></td> 
   <td>对原始Table添加任意列数的任意数值</td> 
   <td>Table, 值为原始数据</td> 
   <td>转换后带有新数列与列名的Table</td> 
   <td></td> 
   <td>Column Expand模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secure_information_retrieval" rel="nofollow">Secure Information Retrieval</a></td> 
   <td></td> 
   <td>通过不经意传输协议安全取回所需数值</td> 
   <td>Table, 值为Instance</td> 
   <td>Table, 值为取回数值</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/transfer_learning/hetero_ftl/README.html" rel="nofollow">Hetero Federated Transfer Learning</a></td> 
   <td>纵向联邦迁移学习</td> 
   <td>在两个party间构建联邦迁移模型</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>FTL神经网络模型参数等</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/psi/README.html" rel="nofollow">PSI</a></td> 
   <td>PSI模块</td> 
   <td>计算两个表特征间的PSI值</td> 
   <td>Table, 值为Instance</td> 
   <td></td> 
   <td></td> 
   <td>PSI 结果</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/unsupervised_learning/kmeans/README.html" rel="nofollow">Hetero KMeans</a></td> 
   <td>纵向 K均值算法</td> 
   <td>构建K均值模块</td> 
   <td>Table, 值为Instance</td> 
   <td>Table, 值为Instance; Arbiter方输出2个Table</td> 
   <td></td> 
   <td>Hetero KMeans模型</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/README.html" rel="nofollow">Data Statistics</a></td> 
   <td>数据统计</td> 
   <td>这个组件会对数据做一些统计工作，包括统计均值，最大最小值，中位数等</td> 
   <td>Table, 值为Instance</td> 
   <td>Table</td> 
   <td></td> 
   <td>Statistic Result</td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/scorecard/README.html" rel="nofollow">Scorecard</a></td> 
   <td>评分卡</td> 
   <td>转换二分类预测分数至信用分</td> 
   <td>Table, 值为二分类预测结果</td> 
   <td>Table, 值为转化后信用分结果</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/util/README.html" rel="nofollow">Sample Weight</a></td> 
   <td>样本加权</td> 
   <td>根据用户设置对输入数据加权</td> 
   <td>Table, 值为Instance</td> 
   <td>Table, 值为加权后Instance</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td><a href="https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/statistic/feldman_verifiable_sum/README.html" rel="nofollow">Feldman Verifiable Sum</a></td> 
   <td>费尔德曼秘密分享求和</td> 
   <td>不暴露隐私数据的前提下进行多方隐私数据求和</td> 
   <td>Table, 值为加数或被加数</td> 
   <td>Table, 值为求和结果</td> 
   <td></td> 
   <td></td> 
  </tr> 
 </tbody> 
</table>

## 安全协议 ##

*  [Encrypt][]
    
     *  [Paillier encryption][]
     *  [Affine Homomorphic Encryption][]
     *  [IterativeAffine Homomorphic Encryption][]
     *  [RSA encryption][]
     *  [Fake encryption][]
 *  [Encode][]
 *  [Diffne Hellman Key Exchange][]
 *  [SecretShare MPC Protocol(SPDZ)][SecretShare MPC Protocol_SPDZ]
 *  [Oblivious Transfer][]
 *  [Feldman Verifiable Secret Sharing][]

[f7ff42452bc71e8f67317e166ba1f6ba.png]: https://img-blog.csdnimg.cn/img_convert/f7ff42452bc71e8f67317e166ba1f6ba.png
[f7ff42452bc71e8f67317e166ba1f6ba.png 1]: https://fate.readthedocs.io/en/latest/_images/federatedml_structure.png
[Encrypt]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#encrypt
[Paillier encryption]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#paillier-encryption
[Affine Homomorphic Encryption]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#affine-homomorphic-encryption
[IterativeAffine Homomorphic Encryption]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#iterativeaffine-homomorphic-encryption
[RSA encryption]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#rst-encryption
[Fake encryption]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#fake-encryption
[Encode]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#encode
[Diffne Hellman Key Exchange]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#diffne-hellman-key-exchange
[SecretShare MPC Protocol_SPDZ]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#secretshare-mpc-protocol-spdz
[Oblivious Transfer]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#oblivious-transfer
[Feldman Verifiable Secret Sharing]: https://fate.readthedocs.io/en/latest/_build_temp/python/federatedml/secureprotol/README.html#feldman-verifiable-secret-sharing