A Privacy Breach Assessment Model for Users Fusing Data Enhancement Policies and BERT Models

In the modern Internet era, the use of online social networks such as Sina Weibo and Facebook is increasing, and at the same time, many users often use them to share their lives or comment on news. According to IDC, unstructured data, such as text data, makes up the majority of the digital space—about 80%—while structured data makes up only 20%. However, this text data may contain sensitive information such as a user's location, educational background, opinions and emotions. Therefore, if this information is utilized by attackers, it will cause serious privacy leakage to users. Therefore, we measure the privacy risk contained in the text data shared by users of online social networks. In this paper, we propose a user Privacy Breach Assessment model that integrates multiple data enhancement strategies and BERT model to effectively alleviate the data imbalance problem of the dataset and improve the assessment accuracy of the model. Specifically, the framework can be divided into three parts: i) the data enhancement part, which consists of multiple data enhancement strategies, including not only the use of EDA methods (synonym substitution, random insertion, random exchange, and random deletion) and back-translation methods to expand the label categories in the dataset with a small number of samples, but also the automated expansion of the data by connecting to the external corpus; ii) a multi-label classification model based on the BERT model; and iii) a multi-label classification model based on the BERT model, which is based on the BERT model and the BERT model. Multi-label classification model based on BERT model, this part mainly utilizes the dataset to fine-tune the training of BERT model, so as to mine the potential privacy leakage of short text data; iii) Calculation and validation of privacy leakage scores, this part mainly involves identifying the historical text data of the social network users through the trained BERT model, and comparing it with the annotation results. The comparison experiments demonstrate that the model proposed in this paper has strong robustness and generalization ability, can effectively handle various types of private text data, and can achieve more accurate classification and prediction in different online social network scenarios. This suggests that the classifier can be used in subsequent privacy risk metrics.