Dataset Properties and Degradation of Machine Learning Accuracy with an Anonymized Training Dataset

With the widespread use of machine learning, we are seeing many use cases where personal data are used for training. There are two privacy threats in such use cases: leakage of training data to a model constructor and inference of personal information from a model by a model user. One countermeasure against these threats is to anonymize training data to preserve privacy. However, it is difficult to find a better anonymization method with less degradation in machine learning performance, such as accuracy, for a given dataset, since anonymization causes inconsistent degradation among datasets. This paper proposes a dataset-property-based method of comparing degradation of machine learning accuracy caused by anonymization. Experiments show that the proposed method can capture the relationship between dataset properties and anonymization methods. They also show the possibility that some dataset properties are related to anonymization and some are not. The proposed method provides a good starting point to help practitioners to choose an anonymization method, that will cause less degradation of machine learning accuracy, for a certain dataset.

I. Introduction

As the use of machine learning becomes widespread, use cases using personal data for training data are also increasing. However, when a model is constructed using personal data in machine learning, this gives rise to two privacy threats. One is data leakage by a model constructor. The constructor can easily obtain personal data since the constructor needs to access the training dataset, including personal data, in order to train a model. The other threat is data leakage by model users. For example, training data can be inferred by model inversion attack [1]. Providing or publicizing the constructed model to others may result in the leaking of information from the training data remaining in the model.