Search

In this paper, we study the problem of privacy-preserving data synthesis (PPDS) for tabular data in a distributed multi-party environment. In a decentralized setting, for PPDS, federated generative models with differential privacy are used by the existing methods. Unfortunately, the existing models apply only to images or text data and not to tabular data. Unlike images, tabular data usually consist of mixed data types (discrete and continuous attributes) and real-world datasets with highly imbalanced data distributions. Existing methods hardly model such scenarios due to the multimodal distributions in the decentralized continuous columns and highly imbalanced categorical attributes of the clients. To solve these problems, we propose a federated generative model for decentralized tabular data synthesis (HT-Fed-GAN). There are three important parts of HT-Fed-GAN: the federated variational Bayesian Gaussian mixture model (Fed-VB-GMM), which is designed to solve the problem of multimodal distributions; federated conditional one-hot encoding with conditional sampling for global categorical attribute representation and rebalancing; and a privacy consumption-based federated conditional GAN for privacy-preserving decentralized data modeling. The experimental results on five real-world datasets show that HT-Fed-GAN obtains the best trade-off between the data utility and privacy level. For the data utility, the tables generated by HT-Fed-GAN are the most statistically similar to the original tables and the evaluation scores show that HT-Fed-GAN outperforms the state-of-the-art model in terms of machine learning tasks. [ABSTRACT FROM AUTHOR]

The threat model of deep-learning-based data theft in data sandbox model was analyzed in detail, and the degree of damage and distinguishing characteristics of this attack were quantitatively evaluated both in the data processing stage and the model training stage. Aiming at the attack in the data processing stage, a data leakage prevention method based on model pruning was proposed to reduce the amount of data leakage while ensuring the availability of the original model. Aiming at the attack in model training stage, an attack detection method based on model parameter analysis was proposed to intercept malicious models and prevent data leakage. These two methods do not need to modify or encrypt data, and do not need to manually analyze the training code of deep learning model, so they can be better applied to data theft defense in data sandbox mode. Experimental evaluation shows that the defense method based on model pruning can reduce 73% of data leakage, and the detection method based on model parameter analysis can effectively identify more than 95% of attacks. [ABSTRACT FROM AUTHOR]