Towards Scalable Privacy-Aware Text Classification: A DistilBERT-Based Approach to PII Identification

For decades, the use of AI chatbots has grown significantly, along with the rise of vulnerabilities, such as the disclosure of confidential information from training data that may include personally identifiable information (PII). Our research aims to further educate users on protecting PII by focusing on detecting the PII risk score based on the assessment of the confidentiality risk within user input. To achieve this purpose, we enhanced the DistilBERT fine-tuned model to recognize each PII class. We took samples from 43.501 sentences containing 56 types of PII data. Based on our samples, we found that FIRSTNAME is the most frequent PII class, with a frequency of 112, followed by EMAIL and LASTNAME. Those samples will be processed through a calculation, resulting in the baseline for three categories of confidentiality classification. Thus, our enhancement will take user input, process it, and produce the result as one of these categories: high, medium, or low. Therefore, the result will provide insight into the use of PII on a user prompt.