Discovering Personally Identifiable Information in Textual Data - A Case Study with Automated Concatenation of Embeddings

Discovering personal identifying information (PII) in unstructured data is an important pre-processing step in enabling privacy preserving machine learning as well as compliance with data protection legislation such as GDPR (General Data Protection Regulation). However, PII discovery in unstructured data, and textual data in particular, is a challenging problem. One of the primary causes is the fact that representations of PII in textual data, do not follow standard grammatical representations. For example, in social media scenarios, textual data can include irregular expressions such as slang and emoticons that must be interpreted contextually to determine whether or not they qualify as PII. The problem is further compounded by the fact that PII discovery algorithms, for the most part, lean on machine learning models that are processing intensive in nature. In scenarios where real-time responses are required to determine whether or not PII exists with in a dataset, this impacts response time negatively. This paper reports on results from a series of experiments in which the Automated Concatenation of Embeddings (ACE) framework was employed to support PII discovery in textual data. Our results show that such architectures that are processing (GPU) intensive are not suitable for handling PII detection in large unstructured or semi-structured data.