Interpretable Privacy Preservation of Text Representations Using Vector Steganography

Contextual word representations generated by language models learn spurious associations present in the training corpora. Adversaries can exploit these to reverse-engineer private attributes of entities mentioned These findings have led efforts towards minimizing privacy risks models. However, existing approaches lack interpretability, compromise on data utility and fail provide guarantees. Thus, goal my doctoral research is develop interpretable preservation text that maximize retention guarantee privacy. To this end, I aim study methods incorporate steganographic modifications within vector geometry obfuscate underlying retain distributional semantic properties learnt during training.