保护数据隐私的深度学习训练数据生成方案

深度学习模型训练存在缺少大量带标签训练数据和数据隐私泄露等问题.为了解决这些问题,借由生成对抗网络可生成大量与真实数据同分布的对抗样本的特点,提出了一个基于条件生成对抗网络的深度学习模型训练数据生成方案.该方案采用条件生成对抗网络生成数据,满足了生成大量带标签训练数据的需求;结合数据变形方法实现数据隐私保护,解决了数据隐私泄露的问题.实验结果表明该方案是高效可行的,而且与其他方案相比,其在数据可用性和保护隐私方面具有优势.

Privacy-preserving deep learning training data generation scheme

The deep learning model training has some problems such as lack of large amount of labeled training data and data privacy leakage. To solve these problems, this paper proposed a conditional generative adversarial network(CGAN) based on deep learning model training data generation scheme. By virtue of the characteristics that CGAN could generate sufficient amount of labeled training data with the same distribution of real data, this scheme used CGAN combined with data morphing to satisfy the requirements of large amount of labeled training data generation and data privacy protection simultaneously. Experimental results show that the scheme is efficient and feasible, and it has advantages in data availability and data privacy preservation when compared with other schemes.