Re-identification of home addresses from spatial locations anonymized by Gaussian skew
© Cassa et al; licensee BioMed Central Ltd. 2008
Received: 25 April 2008
Accepted: 12 August 2008
Published: 12 August 2008
Knowledge of the geographical locations of individuals is fundamental to the practice of spatial epidemiology. One approach to preserving the privacy of individual-level addresses in a data set is to de-identify the data using a non-deterministic blurring algorithm that shifts the geocoded values. We investigate a vulnerability in this approach which enables an adversary to re-identify individuals using multiple anonymized versions of the original data set. If several such versions are available, each can be used to incrementally refine estimates of the original geocoded location.
We produce multiple anonymized data sets using a single set of addresses and then progressively average the anonymized results related to each address, characterizing the steep decline in distance from the re-identified point to the original location, (and the reduction in privacy). With ten anonymized copies of an original data set, we find a substantial decrease in average distance from 0.7 km to 0.2 km between the estimated, re-identified address and the original address. With fifty anonymized copies of an original data set, we find a decrease in average distance from 0.7 km to 0.1 km.
We demonstrate that multiple versions of the same data, each anonymized by non-deterministic Gaussian skew, can be used to ascertain original geographic locations. We explore solutions to this problem that include infrastructure to support the safe disclosure of anonymized medical data to prevent inference or re-identification of original address data, and the use of a Markov-process based algorithm to mitigate this risk.
To develop broadly integrated national healthcare information infrastructure, the utility of sharing personally identifiable data for clinical care, public health and research must always be weighed against privacy concerns. For example, automated outbreak detection systems for surveillance of influenza and bioterrorism, use data from a variety of sources (hospitals, clinics, laboratories) for aggregation, analysis and investigation [1–3]. For the detection of spatial clustering among disease cases, these aggregation systems achieve optimal detection sensitivity and specificity when using the most complete, accurate patient location data .
We have previously described a spatial de-identification algorithm that blurs precise point locations for patients, moving them a randomized distance according to a 2-dimensional Gaussian distribution with variance inversely proportional to the square of the underlying population density . Other spatial anonymization approaches that have been employed include random skews, affine transformations, data aggregation techniques, and the use of software agents to preserve confidentiality [6, 7]. Anonymization of patient address data by reassignment of geographic coordinates allows privacy preservation while sharing data for disease surveillance or biomedical research . As the volume of personally-identifiable health data that is electronically transmitted and published has consistently increased , so has the magnitude of the threat to privacy. Geographical information is particularly identifying; we have demonstrated that it is possible to correctly identify most home addresses even from low resolution point-maps commonly published in journal articles .
Here, we quantitatively demonstrate this vulnerability in two common anonymization approaches. We produce multiple anonymized data sets using a single set of addresses and then progressively average the anonymized results related to each address, characterizing the steep decline in distance of the re-identified point to the original location, (and the reduction in privacy) at each stage. Next, we propose and discuss two solutions to this specific class of vulnerabilities. The first tightly couples anonymization to a distributed health infrastructure that exchanges the data, so that it can control the number of copies distributed to any one party. The second is an extension to the spatial anonymization process employing a Markov process for increasing the anonymity of a 2-dimensional data set anonymized by Gaussian skew.
Re-Identification of points anonymized using Gaussian and randomized skew
Additional information was ascertained from multiple anonymized copies of one original set of point locations, significantly weakening the anonymization used. The average distance to the original addresses after one anonymization pass, which represents the previously described  use of an anonymizing algorithm, was 0.69 km. After each point was inferred using the average of fifty Gaussian skew anonymization passes, the mean distance from the average of all of the anonymized points to the original point in the data set was reduced to 0.1 km.
Similarly, when the anonymizing algorithm is a uniform skew (a random skew that involves moving a point randomly within a square), re-identification attempts using the average of several anonymized data sets also reduced data set privacy markedly. The average distance to the original addresses after one anonymization pass, the traditional use of such algorithms, was set at 0.69 km, to match the level of skew used in the 2-dimensional Gaussian data sets. As in the case of the 2-dimensional Gaussian skew, the average distance to the original point was also reduced to just under 0.1 km after averaging 50 anonymized data sets.
Re-identification of data anonymized with Gaussian and randomized skew
In other words, the x- and y-coordinates are independent Gaussian random variables, each having a standard deviation of σ/. Hence, by taking the average of the anonymization passes, one can obtain the equivalent of a single anonymization pass under a less stringent Gaussian skew anonymization strategy with standard deviation σ/; for 100 passes, reducing the skewing standard deviation along each axis by a factor of 10.
Hence as the number of observations increases, the average of the observations tends to fall nearer to the original point.
It is important to note that while the difference in change to the covariance matrices would appear to make the Gaussian and uniform anonymization skews similar in their ability to protect privacy, this is not necessarily the case. The quantifiable estimate of anonymity, k-anonymity (a metric for data set privacy where k is the number of people among whom a specific individual cannot be distinguished ), that is achieved by each method is different – the Gaussian skew yields higher levels of k-anonymity than the randomized skew does with respect to the average distance moved for each case in a data set. We previously described a method for estimating spatial k-anonymity .
One might fear that an adversary could do even better, devising a novel strategy that uses the sequence L1, L2, ..., L n to get even closer to the original point than a Gaussian with this reduced variance; however, this is not possible without additional data. Stein showed that given n observations of a two-dimensional Gaussian random variable, the most efficient estimator of the mean of the Gaussian is simply the average of the points . Although this seems intuitive for two dimensions, it is surprisingly not the case for three and higher dimensions [12, 13].
Anonymizing within a distributed network or health information exchange
Removing other identifying information from data sets to avoid re-linking
The vulnerability described in this paper relies on the ability to link anonymized data sets together using additional identifiers, or other demographic or clinical data. One possible solution is to swap the addresses in a given data set so that they are effectively unlinked with any unique clinical fields or identifiers, such as a medical record number. This unlinking of spatial data from unique identifiers, however, poses additional challenges: unlinking from any demographic identifiers could reduce the ability to conduct informative disease surveillance, or worse, could make it difficult to actually uncover the addresses of clustered cases when necessary, which is certainly a priority for a public health investigation.
This can be mitigated through the use of randomly generated identifiers for each anonymized instance of a specific record, stored for use in re-linking anonymized data with original data. Additionally, when attempting to determine correlates or predictors of disease, these additional fields may prove important for group stratification. With knowledge of the specific anonymization algorithm and background knowledge such as regional demographic data, it may be possible to further weaken some anonymization algorithms even without repeated attempts.
When considering only two dimensional geographical data, the best way to estimate original locations from several anonymized versions of the same original data set is to average the anonymized longitudes and latitudes. However, there are even more advanced re-identification techniques that can be used to improve the resolution of cases in practice, using data sets with three or more dimensions. If additional fields or identifiers are included in the data set, and those fields are in any way not randomly distributed (anything other than a randomly generated identifier), their presence has the potential to help achieve a higher resolution on the spatial coordinates, even if they do not contain geographical information. This is because there may be additional implicit information linked with spatial addresses in the other dimensions (or fields) that can lend intuition about the distribution of the anonymized spatial coordinates, using approaches that are more advanced than averaging of the fields to estimate the original location .
Increasing anonymity using an algorithm based on a Markov process
While infrastructure for controlled exchange of anonymized health data protects against some vulnerabilities, there are still other methods that could reduce the privacy level of a data set. For example, it is still possible to gain insight into the actual distribution of cases anonymized with knowledge of physical boundaries, highly constrained patient distributions, or other clinical or demographic information about cases. Further study is needed to adequately constrain the anonymized geographical distributions of cases such that this risk is minimized.
In order to protect privacy when using spatial skew algorithms, the number of distinct anonymization results or passes that represent the same data must be controlled. Limiting the generation or disclosure of more than one version will avoid re-identification through averaging. Alternative approaches include integration of anonymization into data provisioning systems to achieve such a restricted data release, or the use of a Markov process to generate multiple anonymized data sets of the same records. These approaches avoid running the algorithm anew with each request, reducing the variation that is at the root of the vulnerability.
Geographical test data sets
Population-adjusted 2-dimensional Gaussian skew
These are simply 1-dimensional Gaussians with means equal to the original coordinates x0 and y0, respectively, and standard deviations σ x and σ y . The parameters σ x and σ y are proportional to the desired level of anonymity k, and are inversely proportional to the population density at . In other words, the greater the anonymity desired, or the lower the underlying population density, the farther points are moved on average.
Re-identification through averaging
We gratefully acknowledge John Brownstein for helpful discussion and support on figures. This work was supported by 1 R01 LM007677 from the National Library of Medicine.
- Mandl KD, Overhage JM, Wagner MM, Lober WB, Sebastiani P, Mostashari F, Pavlin JA, Gesteland PH, Treadwell T, Koski E, Hutwagner L, Buckeridge DL, Aller RD, Grannis S: Implementing syndromic surveillance: a practical guide informed by the early experience. J Am Med Inform Assoc. 2004, 11 (2): 141-150. 10.1197/jamia.M1356. 2003/11/25PubMedPubMed CentralView Article
- Reis BY, Kirby C, Hadden LE, Olson K, McMurry AJ, Daniel JB, Mandl KD: AEGIS: a robust and scalable real-time public health surveillance system. J Am Med Inform Assoc. 2007, 14 (5): 581-588. 10.1197/jamia.M2342. 2007/06/30PubMedPubMed CentralView Article
- Bradley CA, Rolka H, Walker D, Loonsk J: BioSense: implementation of a National Early Event Detection and Situational Awareness System. MMWR Morb Mortal Wkly Rep. 2005, 54 Suppl: 11-19. 2005/09/24PubMed
- Olson KL, Grannis SJ, Mandl KD: Privacy protection versus cluster detection in spatial epidemiology. Am J Public Health. 2006, 96 (11): 2002-2008. 10.2105/AJPH.2005.069526. 2006/10/05PubMedPubMed CentralView Article
- Cassa CA GSJ: A context-sensitive approach to anonymizing spatial surveillance data: impact on outbreak detection. J Am Med Inform Assoc. 2006, 13 (2): 160-165. 10.1197/jamia.M1920.PubMedPubMed CentralView Article
- Armstrong MP, Rushton G, Zimmerman DL: Geographically masking health data to preserve confidentiality. Stat Med. 1999, 18 (5): 497-525. 10.1002/(SICI)1097-0258(19990315)18:5<497::AID-SIM45>3.0.CO;2-#.PubMedView Article
- Kamel Boulos MN, Cai Q, Padget JA, Rushton G: Using software agents to preserve individual health data confidentiality in micro-scale geographical analyses. J Biomed Inform. 2006, 39 (2): 160-170. 10.1016/j.jbi.2005.06.003.PubMedView Article
- Statistics NCH: Health, United States. 2006, p.370, 429, 434-
- Brownstein JS, Cassa CA, Mandl KD: No place to hide--reverse identification of patients from published maps. N Engl J Med. 2006, 355 (16): 1741-1742. 10.1056/NEJMc061891.PubMedView Article
- Bleichenbacher D: Chosen ciphertext attacks against protocols based on the RSA encryption standard PKCS #1. 1998, Crypto '98: 1-12.
- Sweeney L: k-anonymity: A model for protecting privacy. Int J Uncertain Fuzz Int J Uncertain Fuzz. 2002, 10 (5): 557-570.View Article
- Stein C: Inadmissibility of the usual estimator for the mean of a multivariate distribution. 1956, 1: 197-206.
- Baranchik AJ: A family of minimax estimators of the mean of a multivariate normal distribution. Ann Math Statist. 1970, 41: 642-645. 10.1214/aoms/1177697104.View Article
- Cassa CA IK: A software tool for creating simulated outbreaks to benchmark surveillance systems. BMC Med Inform Decis Mak. 2005, 5 (1): 22-28. 10.1186/1472-6947-5-22.PubMedPubMed CentralView Article
- SourceForge: AEGIS-CCT. http://www.sourceforge.net/projects/chipcluster/
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.