Generalised Differential Privacy for Text Document Processing

Abstract

We address the problem of how to "obfuscate" texts by removing stylistic clues which can identify authorship, whilst preserving (as much as possible) the content of the text. In this paper we combine ideas from "generalised differential privacy" and machine learning techniques for text processing to model privacy for text documents. We define a privacy mechanism that operates at the level of text documents represented as "bags-of-words" -these representations are typical in machine learning and contain sufficient information to carry out many kinds of classification tasks including topic identification and authorship attribution (of the original documents). We show that our mechanism satisfies privacy with respect to a metric for semantic similarity, thereby providing a balance between utility, defined by the semantic content of texts, with the obfuscation of stylistic clues. We demonstrate our implementation on a "fan fiction" dataset, confirming that it is indeed possible to disguise writing style effectively whilst preserving enough information and variation for accurate content classification tasks.

Leveraging Hierarchical Representations for Preserving Privacy and Utility in Text

“…Our d χ -privacy algorithm is similar to the model introduced by [29] for privacy preserving text analysis, and [30] for author obfuscation. The algorithms are all analogous to that originally proposed by [13] and we describe it here using the Euclidean distance for word embedding vectors.…”

“…Furthermore, for Euclidean models such as [29], [30], the utility degrades badly as the privacy guarantees increase. This is because the noise injected (line 4 of Alg.…”

“…In this section, we describe how we carried out privacy evaluations using an approach similar to [30].…”

“…1) Task: The task is to carry out author obfuscation against an authorship attribution algorithm. 2) Baselines: Just as in [30], the 'adversarial' attribution algorithm is Koppel's algorithm [40]. Evaluation datasets were selected from the PAN@Clef tasks as follows: [40] to get the baseline.…”

“…Each dataset was then passed through our d χ -privacy algorithm to get a new text output. This was done line by line in a manner similar to [30] i.e. all non stop words were considered.…”