$A^{4}NT$: Author Attribute Anonymity by Adversarial Training of Neural\n Machine Translation

Text-based analysis methods allow to reveal privacy relevant author\nattributes such as gender, age and identify of the text's author. Such methods\ncan compromise the privacy of an anonymous author even when the author tries to\nremove privacy sensitive content. In this paper, we propose an automatic\nmethod, called Adversarial Author Attribute Anonymity Neural Translation\n($A^4NT$), to combat such text-based adversaries. We combine\nsequence-to-sequence language models used in machine translation and generative\nadversarial networks to obfuscate author attributes. Unlike machine translation\ntechniques which need paired data, our method can be trained on unpaired\ncorpora of text containing different authors. Importantly, we propose and\nevaluate techniques to impose constraints on our $A^4NT$ to preserve the\nsemantics of the input text. $A^4NT$ learns to make minimal changes to the\ninput text to successfully fool author attribute classifiers, while aiming to\nmaintain the meaning of the input. We show through experiments on two different\ndatasets and three settings that our proposed method is effective in fooling\nthe author attribute classifiers and thereby improving the anonymity of\nauthors.\n