Preserving Privacy in Analyses of Textual Data Tom Diethe Oluwaseyi Feyisetan Amazon Amazon tdiethe@amazon.com sey@amazon.com Borja Balle Thomas Drake Deep Mind Amazon borja.balle@gmail.com draket@amazon.com ABSTRACT Amazon prides itself on being the most customer-centric company on earth. That means maintaining the highest possible standards of both security and privacy when dealing with customer data. This month, at the ACM Web Search and Data Mining (WSDM) Conference, my colleagues and I will describe a way to protect privacy during large-scale analyses of textual data supplied by cus- tomers. Our method works by, essentially, re-phrasing the customer- supplied text and basing analysis on the new phrasing, rather than on the customers’ own language. CCS CONCEPTS • Security and privacy → Privacy protections; Figure 1: The researchers’ technique adds noise (green) to ACM Reference Format: the embedding of a word (orange) from a textual data set, Tom Diethe, Oluwaseyi Feyisetan, Borja Balle, and Thomas Drake. 2020. producing a new point in the embedding space. Then it finds Preserving Privacy in Analyses of Textual Data. In Proceedings of Workshop the valid embedding nearest that point - in this case, the em- on Privacy in Natural Language Processing (PrivateNLP ’20). Houston, TX, bedding for the word ’mobile’. (STACY REILLY) USA, 3 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn 1 DIFFERENTIAL PRIVACY Questions about data privacy are frequently met with the answer ’It’s anonymized! Identifying features have been scrubbed!’ How- ever, studies such as this one from MIT show that attackers can Differential privacy provides a statistical assurance that the ag- de-anonymize data by correlating it with ’side information’ from gregate figure will not leak information about which individuals other data sources. are in the data set. Say there are two data sets that are identical, Differential privacy [2] is a way to calculate the probability that except that one includes Alice’s data and one doesn’t. Differential analysis of a data set will leak information about any individual in privacy says that, given the result of an analysis - the aggregate that data set. Within the differential-privacy framework, protecting figure - the probabilities that either of the two data sets was the privacy usually means adding noise to a data set, to make data basis of the analysis should be virtually identical. related to specific individuals more difficult to trace. Adding noise Of course, the smaller the data set, the more difficult this stan- often means a loss of accuracy in data analyses, and differential pri- dard is to meet. If the data set contains nine people with 15-minute vacy also provides a way to quantify the trade-off between privacy commutes and one person, Bob, with a two-hour commute, the and accuracy. average commute time is very different for data sets that do and Let’s say that you have a data set of cell phone location traces for do not contain Bob. Someone with side information - that Bob fre- a particular city, and you want to estimate the residents’ average quently posts Instagram photos from a location two hours outside commute time. The data set contains (anonymized) information the city - could easily determine whether Bob is included in the about specific individuals, but the analyst is interested only in an data set. aggregate figure - 37 minutes, say. Adding noise to the data can blur the distinctions between anal- yses performed on slightly different data sets, but it can also reduce Copyright ©2020 for this paper by its authors. Use permitted under Creative Commons the utility of the analyses. A very small data set might require License Attribution 4.0 International (CC BY 4.0). Presented at the PrivateNLP 2020 Workshop on Privacy in Natural Language Processing Colocated with 13th ACM the addition of so much noise that analyses become essentially International WSDM Conference, 2020, in Houston, Texas, USA. meaningless. But the expectation is that as the size of the data set PrivateNLP ’20, February 7, 2020, Houston, TX, USA grows, the trade-off between utility and privacy becomes more © 2020 manageable. 2 PRIVACY IN THE SPACE OF WORD 4 HYPERBOLIC SPACE EMBEDDINGS In November 2019, at the IEEE International Conference on Data In the field of natural-language processing, a word embedding is a Mining (ICDM), we presented a paper [4] that, although it appeared mapping from the space of words into a vector space, i.e., the space first, is in fact a follow-up to our WSDM paper [3]. In that paper, of real numbers. Often, this mapping depends on the frequency we describe an extension of our work on metric differential privacy with which words co-occur with each other, so that related words to hyperbolic space. tend to cluster near each other in the space: So how can we go about preserving privacy in such spaces? One possibility is to modify the original text such that its author cannot be identified, but the semantics are preserved. This means adding noise in the space of word embeddings. The result is sort of like a game of Mad Libs, where certain words are removed from a sentence and replaced with others. While we can apply standard differential privacy in the space of word embeddings, doing so would lead to poor performance. Differential privacy requires that any data point in a data set can be replaced by any other, without an appreciable effect on the results of aggregate analyses. But we want to cast a narrower net, replacing a given data point only with one that lies near it in the semantic space. Hence we consider a more general definition known as ’metric’ differential privacy [1]. Figure 2: A two-dimensional hyperboloid 3 METRIC DIFFERENTIAL PRIVACY I said that differential privacy requires that the probabilities that a The word-embedding space we describe in the WSDM paper is statistic is derived from either of two data sets be virtually identi- the standard Euclidean space. A two-dimensional Euclidean space is cal. But what does ’virtually’ mean? With differential privacy, the a plane. A two-dimensional hyperbolic space, by contrast, is curved. allowable difference between the probabilities is controlled by a In hyperbolic space, as in Euclidean space, distance between parameter, epsilon, which the analyst must determine in advance. embeddings indicates semantic similarity. But hyperbolic spaces With metric differential privacy, the parameter is epsilon times the have an additional degree of representational capacity: the different distance between the two data sets, according to some distance curvature of the space at different locations can indicate where metric: the more similar the data sets are, the harder they must be embeddings fall in a semantic hierarchy [5]. to distinguish. So, for instance, the embeddings of the words ’ibuprofen’, ’medi- Initially, metric differential privacy was an attempt to extend the cation’, and ’drug’ may lie near each other in the space, but their principle of differential privacy to location data. Protecting privacy positions along the curve indicate which of them are more specific means adding noise, but ideally, the noise should be added in a way terms and which more general. This allows us to ensure that we that preserves aggregate statistics. With location data, that means are substituting more general terms for more specific ones, which overwriting particular locations with locations that aren’t too far makes personal data harder to extract. away. Hence the need for a distance metric. In experiments, we applied the same metric-differential-privacy The application to embedded linguistic data should be clear. But framework to hyperbolic spaces that we had applied to Euclidean there’s a subtle difference. With location data, adding noise to a space and observed 20-fold greater guarantees on expected privacy location always produces a valid location - a point somewhere on in the worst case. the earth’s surface. Adding noise to a word embedding produces a new point in the representational space, but it’s probably not the location of a valid word embedding. So once we’ve identified such a point, we perform a search to find the nearest valid embedding. Sometimes the nearest valid embedding will be the original word itself; in that case, the original word is not overwritten. In our paper, we analyze the privacy implications of different choices of epsilon value. In particular, we consider, for a given epsilon value, the likelihood that any given word in a string of words will be overwritten and the number of semantically related words that fall within a fixed distance of each word in the embedding space. This enables us to make some initial arguments about what practical epsilon values might be. Figure 3: A two-dimensional projection of word embeddings in a hyperbolic space. More-general concepts cluster toward the center, more specific concepts toward the edges. 5 BIOGRAPHY Dr. Tom Diethe is an Applied Science Manager in Amazon Research, Cambridge UK. Tom is also an Honorary Research Fellow at the University of Bristol. Tom was formerly a Research Fellow for the “SPHERE” Interdisciplinary Research Collaboration, which is designing a platform for eHealth in a smart-home context. This platform is currently being deployed into homes throughout Bristol. Tom specializes in probabilistic methods for machine learning, applications to digital healthcare, and privacy enhancing technolo- gies. He has a Ph.D. in Machine Learning applied to multivariate signal processing from UCL, and was employed by Microsoft Re- search Cambridge where he co-authored a book titled ‘Model-Based Machine Learning.’ He also has significant industrial experience, with positions at QinetiQ and the British Medical Journal. He is a fellow of the Royal Statistical Society and a member of the IEEE Signal Processing Society. REFERENCES [1] Konstantinos Chatzikokolakis, Miguel E Andrés, Nicolás Emilio Bordenabe, and Catuscia Palamidessi. 2013. Broadening the scope of differential privacy using metrics. In International Symposium on Privacy Enhancing Technologies Sympo- sium. Springer, 82–102. [2] Cynthia Dwork. 2008. Differential privacy: A survey of results. In International conference on theory and applications of models of computation. Springer, 1–19. [3] Oluwaseyi Feyisetan, Borja Balle, Thomas Drake, and Tom Diethe. 2020. Privacy- and Utility- Preserving Textual Analysis via Calibrated Multivariate Perturbations. In Proceedings of the 13th International Conference on Web Search and Data Mining. [4] Oluwaseyi Feyisetan, Tom Diethe, and Thomas Drake. 2019. Leveraging Hi- erarchical Representations for Preserving Privacy and Utility in Text. In IEEE International Conference on Data Mining (ICDM). [5] Maximillian Nickel and Douwe Kiela. 2017. Poincaré embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems. 6338–6347. A version of this first appeared on the Amazon science blog at: https://www.amazon.science/blog/preserving-privacy-in-analyses-of-textual-data