| 差分隐私软大间隔聚类 |
 点此下载全文 |
| 引用本文：谢云轩.差分隐私软大间隔聚类[J].计算技术与自动化,2022,(3):64-70 |
| 摘要点击次数: 183 |
| 全文下载次数: 0 |
|
|
| 中文摘要:软大间隔聚类(Soft Large Margin Clustering)已被证明比其他诸如K-Means等诸多聚类算法具有更优的聚类性能与可解释性。然而作为单机聚类算法，仍有可扩展性的瓶颈，因此有人将其进行分布式改造。然而在进行分布式运算时，在迭代过程中存在节点之间相互通信的过程。如果某些节点存在隐私数据，那么数据集中的敏感信息在通信过程中就可能泄漏。为此，本文将分布式软大间隔聚类算法(Distributed Sparse SLMC)结合隐私保护，通过插入高斯噪声来提供零集中差分隐私（Zero Concentrated Differential Privacy），发展出差分隐私软大间隔聚类算法。最后通过理论证明其隐私保护效用，通过实验验证其具有与非联邦算法相近的收敛速度与聚类性能。 |
| 中文关键词:差分隐私 软大间隔聚类 隐私保护 联邦学习 |
| |
| Differentially Private Soft Large Margin Clustering |
|
|
| Abstract:Soft large margin clustering (SLMC) has been proven to achieve better accuracy and interpretability than other clustering algorithms. However, as a stand-alone clustering algorithm, it still has scalability bottleneck. So researchers develop it into distributed version. However, when performing distributed learning, there are communications between nodes during iterative process. If some nodes have private data, the sensitive information may be leaked in the communication process. So this paper combines the distributed sparse SLMC (DS-SLMC) with privacy protection and provides zero concentration differential privacy by adding Gaussian noise to develop differentially private SLMC (DP-SLMC). Finally, its privacy protection utility is proved theoretically, and experiments show that it has similar convergence speed and clustering performance compared with non-federated algorithm. |
| keywords:differential privacy soft large margin clustering privacy protection federated learning |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
