Privacypreserving record linkage in large databases using secure multiparty computation
 545 Downloads
Abstract
Background
Practical applications for data analysis may require combining multiple databases belonging to different owners, such as health centers. The analysis should be performed without violating privacy of neither the centers themselves, nor the patients whose records these centers store. To avoid biased analysis results, it may be important to remove duplicate records among the centers, so that each patient’s data would be taken into account only once. This task is very closely related to privacypreserving record linkage.
Methods
This paper presents a solution to privacypreserving deduplication among records of several databases using secure multiparty computation. It is build upon one of the fastest practical secure multiparty computation platforms, called Sharemind.
Results
The tests on ca 10 million records of simulated databases with 1000 health centers of 10000 records each show that the computation is feasible in practice. The expected running time of the experiment is ca. 30 min for computing servers connected over 100 Mbit/s WAN, the expected error of the results is 2^{−40}, and no errors have been detected for the particular test set that we used for our benchmarks.
Conclusions
The solution is ready for practical use. It has welldefined security properties, implied by the properties of Sharemind platform. The solution assumes that exact matching of records is required, and a possible future research would be extending it to approximate matching.
Keywords
Secure multiparty computation Privacypreserving record linkage Deduplication PrivacyAbbreviations
 AES
Advanced encryption standard
 CSV
Comma separated values
 PPRL
Privacypreserving record linkage
 SMC
Secure multiparty computation
 SHA
Secure hash algorithm
 UOWHF
Universal oneway hash function
 VM
Virtual machine
 WAN
Wide area network
Background
In this paper, we present a solution to the first track of iDASH 2017 competition [1], titled Deduplication for Global Alliance for Genomics and Health (GA4GH). The goal is to develop privacypreserving record linkage (PPRL) technique on top of existing European ENCCA Unified Patient Identifier (EUPID) framework to facilitate the deduplication task in GA4GH [2].
There is a large number of data providers, called healthcare centers. They all have lists of patients. They want to know whether anyone else has the same patient in the list. This knowledge would allow them to exclude the duplicates from further data analysis, making it more efficient and less biased. Since each center uses its own local identities for its patients, they need to use socalled quasiidentifiers such as patient name and his date of birth to link the records, and this information may be sensitive. The centers could potentially agree on some kind of anonymization, replacing the quasiidentifiers with some random numbers. However, this is in general not sufficient, as e.g. seeing that some patient is present in exactly the centers A, B, C could already potentially leak more information about the patient. For example, if the observer already knew the patient records of A, he would in addition learn that these patients have also visited the centers B and C. Mechanisms stronger than anonymization are needed to overcome this problem.
Technical details of the task
In the competition setting, the patient list of a health center is presented as a table, the columns of which are correspond to the patient’s sensitive data, such as “first name”, “last name”, “date of birth”, “sex”, etc. Two records are considered the same if a certain combination of their attributes matches, e.g. concatenation of the name and the date of birth. For simplicity, it is assumed that the records are already preformatted, and possible typos should not be taken into account (we will discuss exact and approximate matching in the “Discussion” section). For each record, the centers compute the required combination of the attributes, and hash it. They then upload all hashes to the ideal computing machine, implemented with the help of Secure Multiparty Computation (SMC). The machine tells to each center, which of its records are also present in someone else’s dataset, but it does not tell to whom exactly these records belong. The machine itself should be realized by k≥3 specially assigned servers, and no server should learn the values of uploaded hashes, since hashing is used only for shortness of representation and does not provide privacy by itself.
The targeted size of the computation is ca. 1000 centers, with ca. 10000 records each. The competition organizers have made available a test set with 1000 tables, each containing around 10000 records.

The data is not uploaded instantaneously by all centers. It is uploaded over time, so the ideal computing machine may have significant resources left over, while waiting for the next inputs.

The first uploader of each record does not get it indicated as a duplicate, even if more copies of the same record are uploaded later. The record is indicated as a duplicate only to the next uploaders.
 1.
In our first solution, the results are sent out to the centers only after all uploads have finished. It is fine as far as all 1000 centers will agree to contribute their data more or less simultaneously, so that the first uploader will not have to wait for too long. We provide a faster and a more secure algorithm for this setting.
 2.
In our second solution, each upload of a bunch of records by a center is immediately followed by a response from the servers. This algorithm has to be used in the case when some centers may need more time to contribute their data, but the faster uploaders already want to get the results. It is more difficult to preserve privacy in this case.
Related work
There are numerous papers about PPRL, including (but not limited to) solutions based on Bloom filters for string approximate matching [3, 4], extensions of Bloom filters to matching of numerical data [5], as well as provably secure multiparty protocols [6]. Some practical implementations of PPRL include [7], describing a protocol where the linking is delegated to the patient himself, and [8], based on matching seeded hashes, which describes in details which preprocessing steps should be applied to the data to make it better linkable. An extensive stateofthe art overview of this topic in the beginning of 2017 has been given in [9].
Similarly to other privacypreserving computation tasks, the two main approaches to PPRL are Secure Multiparty Computation (SMC) techniques and data perturbation techniques. The SMC approach is in general less efficient, but it provides clear security guarantees, while data perturbation methods can be severely broken due to informal security definitions that they provide [10]. For example, some cryptanalysis of Bloom filter based solutions has been performed in [11, 12]. They show some interesting attacks, including patient name frequency analysis. The iDASH 2017 competition has been focused on SMC methods.
SMC often uses the notion of three separate roles for parties [13]. There are input parties that provide the computation with data, which needs protection. There are computation parties that execute the privacypreserving computation on provided data. The computation parties should not learn anything new from the execution, besides what is leaked through the design of the application. Finally, there are result parties who learn the results of the computation. These three sets of parties may intersect. In the context of the competition task, the computation parties are specially assigned servers, whose number does not depend on the total number of health centers participating. The centers themselves are both input and result parties, each of them receiving a different subset of the outcomes.
 1.
Honest majority: at most 1 of 3 servers can be corrupted.
 2.
Honestbutcurious (passive) adversaries: the corrupted server follows the protocol, but the attacker may get access to all messages received by this server.
As far as these assumptions are satisfied, the only things that may be inferred by the servers from the computation are the following. The first type of leakage is the “duplication pattern”: for each \(n\in \mathbb {N}\), how many records are there, that occur in the lists of all data providers exactly n times? In the context of our task, this leaks how many patients have been registered in exactly n hospitals, but this number is not correlated with particular records or hospitals in any way. The second type of leakage is: how many duplicates have been reported to the health center? It does not allow to correlate the duplicates of different centers in any way, and it only leaks that some n patients of this center have visited some other center. We note that, in our solutions, either the first or the second type of leakage takes place, but never both at once.

Scalability: The network communication complexity of the first algorithm is O(n), where n is the total number of records that all the clients hold. e.g. if there are m clients holding k records each, then n=mk. The complexity of the second algorithm is formally O(n^{2}), but we keep the constant small by using an elaborated filtering technique, which allow to reduce the number of comparisons up to 10 times compared to straightforward pairwise comparison, while slightly sacrificing the answer precision. We note that the running time of the second solution is actually more affected by heavy local computation than by network communication.

Linkage quality: Our solution aims to achieve exact matching, as it was required by the test dataset. The loss in correctness is limited by 2^{−40} in our application. It happens mainly due to the initial application of a hash function to the inputs, which was a competition’s requirement anyway, and also due to dropping some bits in order to match data types that Sharemind supports. In the second solution, some more correctness is sacrificed to gain better efficiency.

Privacy: It is guaranteed that we leak no more than the total number of patients that have been registered in exactly n hospitals (the first solution), or the number of duplicated records in one client’s data (the second solution).
Methods
The Sharemind platform
Our SMC protocols are implemented on top of the Sharemind^{®;} platform [14]. The platform provides a distributed virtual machine (VM) that must be installed at each of the computing servers. The machine interprets the description of a privacypreserving application (in our case, the deduplication), where the cryptographic details are abstracted away. From the application developer’s point of view, different pieces of data are merely labeled as “public” or “private”. The private data will never be learned by any of the servers, unless it is deliberately declassified (i.e. opened to all servers) by the application. While the public values are stored at each server in plain, the private values are stored in a secretshared manner, preventing a single server from learning its value. Underneath, the virtual machine executes cryptographic protocols to perform operations on private secretshared data, which in general requires interaction between the servers. The underlying cryptographic protocols have been proven to be composable [15], meaning that the applications do not need to undergo any additional security proofs. Only deliberate declassification of values needs to be justified. This also concerns the deduplication that we present in this paper.
The main protocol set of Sharemind [16], denoted shared3p, is based on additive sharing among three parties. The private representation of a value u∈R from a finite ringR is denoted by ⟦u⟧, and is defined as ⟦u⟧=(⟦u⟧_{1},⟦u⟧_{2},⟦u⟧_{3}), where the share ⟦u⟧_{i}∈R is held by the ith server. The shares are random elements of R, with the constraint that the sum up to u. Hence, if anyone manages to get only one or two shares of u, he cannot infer any information about u. In [16], the authors have presented protocols for a number of basic arithmetic, relational and logical operations, such as addition, multiplication, comparisons, etc., transforming shares of the inputs into shares of the output. The supported rings are \(\mathbb {Z}_{2^{n}}\) and \(\mathbb {Z}_{2}^{n}\), and several different rings may be in used simultaneously in the same application. There are special protocols for converting shares between different rings. This basic set of operations has been extended in numerous followup papers [17, 18, 19, 20, 21, 22].
The protocols of shared3p are secure against one passively corrupted party. Our deduplication application is built on top of them. In the rest of this section, we present all algorithms that we used in our application. In the loops, we use foreach and while to denote parallelizable loops, whilst for loops are sequential. Parallelization is important due to the latency of all nonlinear operations with private values, due to the need to exchange messages between computation servers. We denote vectors as \(\vec {x} = \langle {x_{1},\ldots,x_{n}}\rangle \).
The first solution: output all results once in the end
In out first solution, the results of computation are only available at the end, after all clients (the health centers) have uploaded the hashes of their records. The outline of the process is the following. In the first phase, the SMC servers are collecting input data from the clients, without performing any deduplication detection on it. When the uploads have ceased (e.g. the deadline or some other trigger event has happened), the servers stop collecting the data. They run the deduplication algorithm on all data they managed to collect so far, and give to each client its personal output. If a record is duplicated, then the client that first uploaded it will not be notified that it is a duplicate. All other clients that have uploaded the same record will receive a notification about it.
Cryptographic building blocks
We fix two cryptographic functions: the hash function H:{0,1}^{∗}→{0,1}^{η}, and a block cipher \(E:\{0,1\}^{\eta }\times \mathcal {K}\rightarrow \{0,1\}^{\eta }\), where \(\mathcal {K}\) is the set of possible keys for E. The block size of E should be the same as the output length of H, so that we could apply E directly to the output of H. The challenge is that we will apply E to secretshared values, and hence the block cipher E has to be easily computable under SMC. We have picked AES128 as E, and the privacypreserving implementation of AES128 is already available in Sharemind [19]. There exist newer, possibly more efficient implementations of AES [23], as well as proposals for SMCoptimized block ciphers [24], which have not been implemented on Sharemind yet, and could potentially speed the computation up.
In our solution, we have taken η=128. We let H be the composition of SHA256 cryptographic hash function and a universal oneway hash function family (UOWHF) [25].
Computation on the client side
The behaviour of a client is described in Algorithm 1. At startup, each client queries the Sharemind servers for the random parameters of the UOWHF H, which the servers generate in the beginning. Each client takes the input records as soon as they become available, picks the necessary attributes from each record, applies H to them, and uploads the hashes to Sharemind servers in secretshared manner. The Sharemind API readily supports that. At this point, no server learns the exact values of the hashes, since they are all secretshared.
In the end, when the servers have finished the computation, the client queries them for its personal result. The servers respond with the shares of the output, which is a vector of booleans of the same length as the number of records from this client, indicating whether a record is a duplicate one. The client reconstructs the result vector. Again, the Sharemind API already has support for that.
Computation on the server side
We describe the work of the servers in phases.
Startup.

parameters of the hash function H;

a key ⟦K⟧ for the block cipher E.
All these values are generated in such a way that they remain secretshared among the servers, and no server actually learns them. Sharemind API supports such shared randomness generation, and we denote the corresponding functionality as random.
The servers initialize a public variable cnt ←0. It will be used for indexation of clients.
Upload
During the upload phase, two different activity threads are essentially taking place in parallel.
One thread is the actual acceptance of data from the clients, described in Algorithm 3. The hashes of clients are stored into an private array ⟦v⟧, and the corresponding client identities into a public array s under the same indices. In principle, this algorithm could be invoked several times for the same client, if it intends to split up the upload of its data. Several clients may want to connect to the servers at the same time, so we need to make use of Sharemind’s database support to avoid race condition on cnt.
The other thread of activity is encrypting the elements ⟦v_{i}⟧. For each i, the servers evaluate E on ⟦v_{i}⟧ in a privacypreserving way, using the same key ⟦K⟧ for each encryption, and take the first 64 bits of the result, as described in Algorithm 4. The only reason why we take only 64 of 128 bits is that the largest ring that Sharemind currently supports is \(\mathbb {Z}_{2^{64}}\). There are no problems with taking only 64 bits, as long as there are no collisions. With the envisioned amounts of data (around 10^{7}=2^{23.5} records in total), collisions are unlikely; the birthday paradox puts their probability at around 2^{−64+2·23.5}=2^{−17}. We note that the communication complexity also reduces with the number of bits of the data types.
In our implementation, evaluation of E can take place either immediately after the data upload, or it can be run in a separate thread, waiting until more data comes to process it all in parallel. In any case, the application of E is done in parallelized manner, batching up the hashes that are waiting for being processed so far.
Computation of results.
When the upload has finished and the ciphertext ⟦c_{i}⟧ has been computed for each i∈{0,…,cnt−1}, then the final results are computed as in Algorithm 5. First of all, all computed ciphertexts are privately shuffled, and then declassified. Since they are encrypted with a block cipher whose key remains secretshared, it only reveals the “duplication pattern”, i.e. how many values occur there exactly n times. The private shuffle [26] in the shared3p protocol set is highly efficient: the amount of communication is linear in the size of the shuffled vector, and the number of communication rounds is constant. It also allows to easily apply the shuffle inverse, which is as efficient as the shuffle itself.
At this point, the servers could already mark the duplicates with boolean values 1 and the remaining values with 0. After the obtained vector is privately shuffled back, the servers can return the secret shared bits to the clients, so that the ith clients learns the bits indicating the duplicity of its own records. The problem is that the servers need to notify all clients except the first one, but because of shuffling they do not know which entry belongs to the first client. They cannot declassify the client indices ⟦s_{i}⟧ either since it would partially undo the shuffling.
In order to determine which value belongs to the first client, the servers run Algorithm 6. This algorithm is applied to each set of shuffled entries that have identical values, to determine the minimum amongst them. The minimum should be labeled false since it is not considered a duplicate, and all other elements should be labeled true. The idea behind this recursive algorithm is the following. The inputs are split into pairs, and a comparison performed within each pair. The element that is larger is definitely not the minimum of the entire set, so we can immediately write b_{i}:=true. The indices of all elements that turned out to be smaller are stored into m_{i}, and the whole algorithm is now applied again to the elements indexed by m_{i}. The procedure is repeated until there are is only one element left, which is the minimum, so the algorithm returns false in oneelement case.
Although Algorithm 6 works with private values, it reveals the ordering between certain elements of \(\vec {t}\). Because of the random shuffle, the ordering of this vector is completely random, so it does not disclose any information. No server learns from the output more than it would from a random permutation.
The second solution: output results immediately
Our second solution considers the case where the client is immediately notified, which of its hashes of records have already been uploaded by some previous client. This task is more difficult than the previous one. We cannot simply declassify the encrypted hashes when they arrive, since it would leak which pairs of centers have intersecting sets of records.
The second solution has the same components E and H as the first solution, constructed in the same manner. Again, the computation servers agree on the key ⟦K⟧ in the beginning, as well as the parameters of H.
Computation on the client side
The clients perform pretty much in the same manner as in the first solution, according to Algorithm 1. The only technical difference is that the servers react to the upload immediately, so the client most likely does not interrupt the session with the servers, and stays connected until it receives the result.
Computation on the server side
Assume that the servers already store T encrypted hashes (zero or more), provided by the previous clients. These T hashes have been stored as public values z_{1},…,z_{T}, where z_{i} is equal to the first 64 bits of E(K,v_{i}). Duplicates have already been removed among z_{1},…,z_{T}. Assuming that duplicates never end up among z_{1},…,z_{T}, since ⟦K⟧ remains private, z_{1},…,z_{T} are computationally indistinguishable from T uniformly randomly distributed values, so it is safe to make them public.
The 64bit values z_{1},…,z_{T} are kept in 2^{B}buckets, numbered from 0 to 2^{B}−1. In our implementation, B=16. Each bucket is just a vector of values. Each z_{i} is assigned to the bucket B_{j}, where j is equal to the first 16 bits of z_{i}.
Now suppose that a center has uploaded the hashes ⟦v_{1}⟧,…,⟦v_{t}⟧. The servers need to check whether these hashes has occurred before. The computation starts by encrypting each ⟦v_{i}⟧. Let \(\llbracket {z^{\prime }_{1}}\rrbracket,\ldots,\llbracket {z^{\prime }_{t}}\rrbracket \) be the results of encryption, computed by \(\llbracket {z^{\prime }_{i}}\rrbracket =\mathsf {{leftbits}_{64}}(E(\llbracket {K}\rrbracket,\llbracket {v_{i}}\rrbracket))\). We cannot immediately declassify them since it would leak which pairs of centers have intersecting sets of records. Instead, we should use privacypreserving comparison.
We do not want to simply invoke the private comparison protocol for each z_{i} and \(\llbracket {z^{\prime }_{j}}\rrbracket \), because we consider their number to be too large. Indeed, as we are aiming to handle ca. 10 million records, this method would cause us to compare each pair of records, leading to ca. 5·10^{13} invocations of the comparison protocol. An ℓbit comparison requires slightly more network communication than a ℓbit multiplication [16], with the latter requiring each computation server to send and receive two ℓbit values [17]. If ℓ=64 and there are 5·10^{13} operations, then each server has to send out and receive at least 6·10^{15} bits, which on a 100 Mb/s network (specified in the conditions of the competition task) would take almost two years.
We reduce the number of comparisons in the following manner. Let ⟦z^{′}⟧ be one of the private values \(\llbracket {z^{\prime }_{1}}\rrbracket,\ldots,\llbracket {z^{\prime }_{t}}\rrbracket \); all t values are handled in parallel. The comparison of ⟦z^{′}⟧ with z_{1},…,z_{T} is described in Algorithm 7. In this algorithm, we let N be the maximum size of a bucket. We denote the jth element of the ith bucket by B_{i,j}. We assume that each bucket has exactly N elements, adding special dummy elements if necessary.
The characteristic vector of an element \(x\in \mathbb {Z}_{M}\) is a vector \(\langle {b_{0},\ldots,b_{M1}}\rangle \in \mathbb {Z}_{2}^{M}\), where b_{x}=1 and all other elements are equal to 0. The shared3p protocol set has a simple and efficient protocol for computing characteristic vectors, described in [27]. The protocol turns a private value into a private characteristic vector. The characteristic vector of leftbits_{16}(⟦z^{′}⟧) marks the index of the bucket to which ⟦z^{′}⟧ belongs, and the expression \(\bigoplus _{i=0}^{2^{B}1}\llbracket {b_{i}}\rrbracket \cdot \mathbf {B}_{i,j}\) returns exactly the jth element of that bucket, which we denote ⟦y_{j}⟧. This way, the values ⟦y_{1}⟧,…,⟦y_{N}⟧ are the privately represented content of the bucket, into which z^{′} would belong. The private comparison \(\llbracket {z^{\prime }}\rrbracket \stackrel {?}{=}\llbracket {y_{j}}\rrbracket \) is performed for all j, thus comparing ⟦z^{′}⟧ against each element that belongs to the ith bucket. Finally, \(\llbracket {b}\rrbracket =\bigvee _{j=1}^{n}\llbracket {c_{j}}\rrbracket \) is the private OR of all comparisons, which tells whether there had been at least one match. The private bits ⟦b_{1}⟧,…,⟦b_{t}⟧ resulting from applying Algorithm 7 to all \(\llbracket {z^{\prime }_{1}}\rrbracket,\ldots,\llbracket {z^{\prime }_{t}}\rrbracket \) are returned to the client.
 1.
randomly shuffle 〈⟦b_{1}⟧,…,⟦b_{t}⟧〉 and \(\langle {\llbracket {z^{\prime }_{1}}\rrbracket,\ldots,\llbracket {z^{\prime }_{t}}\rrbracket }\rangle \), using the same permutation;
 2.
declassify \(\llbracket {\vec {b}}\rrbracket \);
 3.
declassify those ⟦zi′⟧, where b_{i}=0, and add these \(z^{\prime }_{i}\) to the respective buckets.
In our implementation, ⟦z^{′}⟧ is shared over \(\mathbb {Z}_{2}^{64}\), hence taking the first 16 bits of it is a local operation, resulting in a value in \(\mathbb {Z}_{2}^{16}\). The characteristic vector protocol in [27] is easily adaptable to compute the characteristic vectors of elements of \(\mathbb {Z}_{2}^{B}\), and its result is a vector over \(\mathbb {Z}_{2}\) with length 2^{B}. The computation of the characteristic vector requires communication of two elements of \(\mathbb {Z}_{2}^{B}\) and one element of \(\mathbb {Z}_{2}^{2^{B}}\) (in total, not per party).
The computation of ⟦y_{j}⟧ in Algorithm 7 is again a local operation. The computations of ⟦c_{j}⟧ and their disjunction are straightforward using the protocols available in the shared3p set of Sharemind.
Speeding up local computations
In Alg. 7, the computation of ⟦y_{j}⟧ is local. Nevertheless, in practice it takes a major part of the entire effort of the protocol, taking up most of the time in it. If we knew something about the size of z^{′}, we could compute \(\bigoplus \) not over all buckets, but only over a subset of them, covering only the range into which z^{′} is guaranteed to fall, with a negligible error.
A bucket is defined by the most significant bits of elements that it contains. If we take any two buckets, then all elements in one of them will be strictly smaller than all elements in the other one. If we sort \(\llbracket {z^{\prime }_{1}}\rrbracket,\ldots,\llbracket {z^{\prime }_{t}}\rrbracket \) in ascending order (Sharemind has efficient protocols for sorting private values [28]), we know that the first elements more likely belong to the buckets with “smaller” bits, and the last elements more likely belong to buckets with “larger” bits. We can estimate these probabilities more precisely.
As the key ⟦K⟧ is secret, and the hashes ⟦v_{1}⟧,…,⟦v_{t}⟧ are all different, the values \(\llbracket {z^{\prime }_{1}}\rrbracket,\ldots,\llbracket {z^{\prime }_{t}}\rrbracket \) can be treated as mutually independent, uniformly random elements of \(\mathbb {Z}_{2}^{64}\). After sorting them, their likely ranges can be derived from the order statistics as follows.
Let \(\mathcal {P}\) be a discrete probability distribution over values x_{1},x_{2},…, such that the probability mass of x_{i} is p_{i}. Let X_{1},…,X_{n} be random variables sampled from \(\mathcal {P}\), and let \(X^{\prime }_{1},\ldots,X^{\prime }_{n}\) be obtained after sorting X_{1},…,X_{n} in ascending order. We have \(\text {Pr}[X^{\prime }_{j} \leq x_{i}] = \sum _{k=j}^{n}{{n}\choose {k}}P_{i}^{k} \cdot (1  P_{i})^{nk}\), where \(P_{i} = Pr[X_{i} \leq x_{i}] = \sum _{k=1}^{i} p_{k}\). This quantity comes from summing up probabilities of all possible combinations, where at least j of n variables are smaller than x_{i}. For a fixed j, this expression is actually the cumulative density function of a binomial distribution B(n,P_{j}).
By default, we use ε=2^{−40} as the probability of error. As the total number of hashes is expected to be around 1000·10000≈2^{23.5} and we have two bounds to try for each hash, the probability of making a bounds check error during the whole run is not more than 2·2^{−40+23.5}=2^{−15.5}, which we consider acceptable, and which is also similar to errors due to the collisions in the first 64 bits of AES output.
The usefulness of these bounds increases together with t. If t=100 (and ε=2^{−40}), then we gain little, as the ranges [m_{i},M_{i}] still cover around half of the whole range. If t=10000, then the sorted values can be much more tightly positioned — the ranges [m_{i},M_{i}] are less than 1/10 of the whole range.
Alternative comparison
The communication costs of Algorithm 7 may be further reduced, ultimately turning them into a constant (assuming that B is constant), albeit with a further increase in the costs of local computation.
Consider the bucket B_{i} with elements B_{i,1},…,B_{i,N}. The value z^{′} is an element B_{i} iff it is a root of the polynomial \(\mathbf {P}_{i}(x)=\prod _{j=1}^{N}(x\mathbf {B}_{i,j})\). The polynomial is considered over the field \(\mathbb {F}_{2^{64}}\). The elements of this field are 64bit strings and their addition is bitwise exclusive or. Hence, an additive sharing over \(\mathbb {F}_{2^{64}}\) is at the same time also sharing over \(\mathbb {Z}_{2}^{64}\) and vice versa.
Let P_{i,0},…,P_{i,N} be the coefficients of the polynomial P_{i}. It does not make sense to compute P_{i}(⟦z^{′}⟧) in a straightforward way, because this would involve N−1 private multiplications for each bucket. A better way is given in Algorithm 9.
In Algorithm 9, ⟦d_{i}⟧ is computed as a scalar product of the private vector of the powers of z^{′} with the public vector of the coefficients of P_{i}. The powers of ⟦z^{′}⟧ have to be computed only once for all buckets. These powers are computed with the help of the usual multiplication protocol of Sharemind, but working over \(\mathbb {F}_{2^{64}}\). The local operations in this protocol are multiplications in \(\mathbb {F}_{2^{64}}\), which are relatively more expensive than ordinary bitwise operations. In our implementation we use the NTL library [29] for binary field operations. The computation of all ⟦d_{i}⟧ involves many multiplications in this field, so even though the computation does not require any communication between the parties, it is quite heavy on the local side.
As the computation of ⟦z^{′}^{2}⟧,…,⟦z^{′}^{N}⟧ is done over a field, we can push its heaviest part to the precomputation phase, leaving just a single private multiplication to be done during runtime. The method is described in ([30], Algorithm 1). The precomputation consists of generating a random invertible element \(\llbracket {r}\rrbracket \in \mathbb {F}_{2^{64}}^{*}\) together with its inverse ⟦r^{−1}⟧ and computing ⟦r^{2}⟧,…,⟦r^{N}⟧. During the runtime, one computes ⟦z^{′}⟧·⟦r^{−1}⟧ and declassifies it. For an exponent k, the private value ⟦z^{′}^{k}⟧ is then found as ⟦z^{′}^{k}⟧=(z^{′}·r^{−1})^{k}·⟦r^{k}⟧, which can be computed locally, without interaction.
In binary fields, including \(\mathbb {F}_{2^{64}}\), squaring of additively shared values does not require communication between the servers. This can be used to speed up precomputations. In ([30], Algorithm 4) it is shown how to reduce the communication cost of computing ⟦r^{2}⟧,…,⟦r^{N}⟧ to that of approximately \(\sqrt {N}\) multiplications.
The polynomialbased comparison method is also amenable to the order statistic related speedup described in the previous section. Both comparison methods have been implemented in our second solution.
Results
We have evaluated our solution on the test data provided by iDASH 2017 competition organizers. The data consists of 1000 CSV formatted files, ca. 10000 rows each. The evaluation has been performed on a local cluster of three 2 x Intel Xeon E52640 v3 2.6 GHz/8GT/20M servers, where one server has also played the role of all clients. The network bandwidth has been throttled to 100 Mb/s, and latency changed to 40 ms. The number of computation threads was limited to 2, as required by the competition rules. The data upload times have not been measured since they depend on the network connection between the server and the centers, and are not related to the particular SMC algorithms.
The number of incorrect answers in our benchmarks was 0, and it turns out that the error 2^{−40} was fine for 1000 clients with 10000 records each.
First solution
First solution
Encrypting the hashes  29 m 
Shuffling the encrypted hashes  60 s 
Find the first element in each duplicated set  30 s 
Unshuffling the boolean results  12 s 
Public sort, housekeeping  18 s 
Total  32 m 
Second solution
Second solution (w/o polynomials)
Housekeeping (avg per client)  0.28 s 
Encrypting the hashes (avg per client)  6.3s 
Shuffling the encrypted hashes (avg per client)  0.51 s 
Detecting duplicates (min of 1000 clients)  20s 
Detecting duplicates (max of 1000 clients)  46s 
Total (1000 clients)  12 h 45 m 
Second solution (w/ polynomials)
Housekeeping (avg per client)  0.28 s 
Encrypting the hashes (avg per client)  5.8 s 
Shuffling the encrypted hashes (avg per client)  0.53 s 
Detecting duplicates (min of 1000 clients)  14 s 
Detecting duplicates (max of 1000 clients)  666 s 
Total (1000 clients)  88 h 8 m 
Second solution (w/o polynomials, no thread limit)
Housekeeping (avg per client)  0.28 s 
Encrypting the hashes (avg per client)  6.3 s 
Shuffling the encrypted hashes (avg per client)  0.50 s 
Detecting duplicates (min of 1000 clients)  20 s 
Detecting duplicates (max of 1000 clients)  40 s 
Total (1000 clients)  11 h 20 m 
Second solution (w/ polynomials, no thread limit)
Housekeeping (avg per client)  0.28 s 
Encrypting the hashes (avg per client)  5.8 s 
Shuffling the encrypted hashes (avg per client)  0.55 s 
Detecting duplicates (min of 1000 clients)  14 s 
Detecting duplicates (max of 1000 clients)  112 s 
Total (1000 clients)  19 h 30 m 
Discussion
We think that the results are good enough and can be applied to real data. Depending on the requirements, either the first or the second solution can be used.
The data that was provided by iDASH 2017 competition organizers was idealized, i.e. it was already preprocessed and assumed that e.g. the gender is labeled exactly “M” and “F” and there are no other possible encodings. In a practical application, we should think how much preprocessing we agree to delegate to the data owner, and how much we perform using secure multiparty computation.
We note that revealing “duplication pattern” is safe as far as there are sufficiently many clients, so that leaking how many patients have been present in exactly n centers would be safe. This leakage would be bad in some extreme cases, e.g with two medical centers with one patient, since then the server will know whether that particular patient has been in both centers or only one of them. However, if the data is small, then slower and more secure algorithms could be used for that anyway, such as finding privacypreserving set intersection between each pair of medical centers to determine the duplicates.
One possible problem of Sharemind deployment is that its protocols work under assumptions of the honest majority and an honestbutcurious adversary. This means that one needs to find three hosts who agree to set up an SMC server, and who are trustworthy enough not to deliberately break the computation, and clean the server from all intermediate data after the computation has finished (some challenges of realworld application deployment are described e.g. in [31]). Sacrificing in efficiency, we could use protocol sets like SPDZ [32] or MASCOT [33], which are actively secure against a dishonest majority. In any case, as far as we use secret sharing, we need to assume that at least one server should remain uncorrupted, or otherwise the computation parties will collaborate to reconstruct the clients’ secrets, regardless of the privacypreserving algorithms used.
Possible future work
As discussed e.g. in [9], the quasiidentifiers are often inconsistent in real applications, so exact comparison of values is not sufficient to achieve accurate linkage results, and some approximated string matching can be preferred. Our algorithms cannot be extended to this setting directly, since AES encryption maps similar strings to completely different values. We could take some ideas from some previous works, and hash all ngrams of a string instead of the entire string. Different ngrams should be encrypted with different AES keys to avoid leaking similarities in the string patterns. Nevertheless, this would already leak more information, and a corrupted server would be able to perform e.g. analysis of bigram frequencies, and there could possibly be more attacks inspired by [12], even though the attacks are adapted to Bloom filters that we do not use in our solution. We conclude that the extension to approximate matching would require a more careful design and security analysis.
Conclusion
In this paper, we have proposed two methods for finding duplicates between several databases. The implementation has been benchmarked on data of 1000 medical centers with 10000 records each, giving quite positive results, which would be acceptable in realworld scenarios.
Notes
Acknowledgements
The authors want to thank Marju Ignatjeva who helped to set up the Sharemind virtual machine for iDASH competition.
Funding
The research and the publication cost were funded by Estonian Research Council, grant no. IUT271.
Availability of data and materials
The test data itself is not available anymore on the competition site. The virtual machine of the implementation is not available due to Sharemind group’s commercial interests.
About this Supplement
This article has been published as part of?BMC Medical Genomics?Volume 11 Supplement 4, 2018: Proceedings of the 6th iDASH Privacy and Security Workshop 2017. The full contents of the supplement are available online at?https://bmcmedgenomics.biomedcentral.com/articles/supplements/volume11supplement4.
Authors’ contributions
PL has developed the theory part and the corresponding algorithms. AP has implemented these algorithms on Sharemind platform, and performed the experiments. Both authors have read and approved the final manuscripct.
Ethics approval and consent to participate
Not relevant.
Consent for publication
Not relevant
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
 1.iDASH 2017 competition. http://www.humangenomeprivacy.org/2017/. Accessed 7 July 2018.
 2.Ebner H, Hayn D, Falgenhauer M, Nitzlnader M, Schleiermacher G, Haupt R, Erminio G, Defferrari R, Mazzocco K, Kohler J, Tonini G, Ladenstein R, Schreier G. Piloting the european unified patient identity management (eupid) concept to facilitate secondary use of neuroblastoma data from clinical trials and biobanking, vol. 223. In: Health Informatics Meets eHealth: Predictive Modeling in Healthcare  From Prediction to Prevention, Proceedings of the 10th eHealth 2016 Conference, Studies in Health Technology and Informatics. Vienna: IOS Press: 2016. p. 31–38. https://doi.org/10.3233/978161499645331.Google Scholar
 3.Schnell R, Bachteler T, Reiher J. Privacypreserving record linkage using bloom filters. BMC Med Inform Decis Making. 2009; 9(1):41. https://doi.org/10.1186/14726947941.CrossRefGoogle Scholar
 4.Karapiperis D, Verykios VS. A fast and efficient hamming lshbased scheme for accurate linkage. Knowl Inf Syst. 2016; 49(3):861–84. https://doi.org/10.1007/s101150160919y.CrossRefGoogle Scholar
 5.Vatsalan D, Christen P. Privacypreserving matching of similar patients. J Biomed Inform. 2016; 59:285–98. https://doi.org/10.1016/j.jbi.2015.12.004.CrossRefGoogle Scholar
 6.Du W, Atallah MJ. Protocols for secure remote database access with approximate matching, vol. 2 In: Ghosh AK, editor. ECommerce Security and Privacy. Advances in Information Security. Boston: Springer: 2001. p. 87–111. https://doi.org/10.1007/9781461514671.Google Scholar
 7.Alhaqbani B, Fidge C. Privacypreserving electronic health record linkage using pseudonym identifiers. In: HealthCom 2008  10th International Conference on Ehealth Networking, Applications and Services. Singapore: IEEE: 2008. p. 108–17. https://doi.org/10.1109/HEALTH.2008.4600120.Google Scholar
 8.Kho AN, Cashy JP, Jackson KL, Pah AR, Goel S, Boehnke J, Humphries JE, Kominers SD, Hota BN, Sims SA, Malin B, French DD, Walunas TL, Meltzer DO, Kaleba EO, Jones RC, Galanter WL. Design and implementation of a privacy preserving electronic health record linkage tool in chicago. J Am Med Inform Assoc. 2015; 22(5):1072–80. https://doi.org/10.1093/jamia/ocv038.CrossRefGoogle Scholar
 9.Vatsalan D, Sehili Z, Christen P, Rahm E. Privacypreserving record linkage for big data: Current approaches and research challenges In: Zomaya AY, Sakr S, editors. Handbook of Big Data Technologies. Cham: Springer: 2017. p. 851–95. https://doi.org/10.1007/9783319493404.Google Scholar
 10.Kargupta H, Datta S, Wang Q, Sivakumar K. Randomdata perturbation techniques and privacypreserving data mining. Knowl Inf Syst. 2005; 7(4):387–414. https://doi.org/10.1007/s1011500401736.CrossRefGoogle Scholar
 11.Kuzu M, Kantarcioglu M, Durham EA, Toth C, Malin B. A practical approach to achieve private medical record linkage in light of public resources. J Am Med Inform Assoc. 2013; 20(2):285–92. https://doi.org/10.1136/amiajnl2012000917.CrossRefGoogle Scholar
 12.Niedermeyer F, Steinmetzer S, Kroll M, Schnell R. Cryptanalysis of Basic Bloom Filters Used for Privacy Preserving Record Linkage. J Priv Confidentiality. 2014; 6(2):59–79.Google Scholar
 13.Kamm L. Privacypreserving statistical analysis using secure multiparty computation. 2015. PhD thesis, University of Tartu.Google Scholar
 14.Sharmeind secure multiparty computation platform. https://sharemind.cyber.ee. Accessed 7 July 2018.
 15.Bogdanov D, Laud P, Laur S, Pullonen P. From Input Private to Universally Composable Secure Multiparty Computation Primitives In: Datta A, Fournet C, editors. IEEE 27th Computer Security Foundations Symposium, CSF 2014, Vienna, Austria, 1922 July, 2014. Computer Society Press: IEEE: 2014. p. 184–98. https://doi.org/10.1109/CSF.2014.21.Google Scholar
 16.Bogdanov D, Niitsoo M, Toft T, Willemson J. Highperformance secure multiparty computation for data mining applications. Int J Inf Sec. 2012; 11(6):403–18.CrossRefGoogle Scholar
 17.Randmets J. Programming languages for secure multiparty computation application development. 2017. PhD thesis, University of Tartu. http://hdl.handle.net/10062/56298.
 18.Dimitrov V, Kerik L, Krips T, Randmets J, Willemson J. Alternative implementations of secure real numbers. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security. CCS ’16. New York: ACM: 2016. p. 553–64. https://doi.org/10.1145/2976749.2978348.Google Scholar
 19.Laur S, Talviste R, Willemson J. From Oblivious AES to Efficient and Secure Database Join in the Multiparty Setting, vol. 7954. In: Applied Cryptography and Network Security. LNCS. Berlin: Springer: 2013. p. 84–101.Google Scholar
 20.Krips T, Willemson J. Hybrid model of fixed and floating point numbers in secure multiparty computations, vol. 8783. In: Proceedings of the 17th International Information Security Conference, ISC 2014. LNCS. Cham: Springer: 2014. p. 179–97.Google Scholar
 21.Kamm L, Willemson J. Secure floatingpoint arithmetic and private satellite collision analysis. Int J Inf Secur. 2015; 14:531–48.CrossRefGoogle Scholar
 22.Krips T, Willemson J. Pointcounting method for embarrassingly parallel evaluation in secure computation, vol. 9482 In: GarciaAlfaro J, Kranakis E, Bonfante G, editors. Foundations and Practice of Security: 8th International Symposium, FPS 2015, ClermontFerrand, France, October 2628, 2015, Revised Selected Papers. LNCS. Cham: Springer: 2016. p. 66–82.Google Scholar
 23.Araki T, Furukawa J, Lindell Y, Nof A, Ohara K. Highthroughput semihonest secure threeparty computation with an honest majority. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security. CCS ’16. New York: ACM: 2016. p. 805–17. https://doi.org/10.1145/2976749.2978331.Google Scholar
 24.Albrecht MR, Grassi L, Rechberger C, Roy A, Tiessen T. Mimc: Efficient encryption and cryptographic hashing with minimal multiplicative complexity, vol. 10031 In: Cheon JH, Takagi T, editors. Advances in Cryptology  ASIACRYPT 2016  22nd International Conference on the Theory and Application of Cryptology and Information Security, Hanoi, Vietnam, December 48, 2016, Proceedings, Part I. Lecture Notes in Computer Science: 2016. p. 191–219. https://doi.org/10.1007/9783662538876.Google Scholar
 25.Naor M, Yung M. Universal oneway hash functions and their cryptographic applications In: Johnson DS, editor. Proceedings of the 21st Annual ACM Symposium on Theory of Computing, May 1417, 1989, Seattle, Washigton, USA. New York: ACM: 1989. p. 33–43. https://doi.org/10.1145/73007.73011.Google Scholar
 26.Laur S, Willemson J, Zhang B. RoundEfficient Oblivious Database Manipulation. In: Proceedings of the 14th International Conference on Information Security. ISC’11, vol. 7001. Lecture Notes in Computer Science. Xi’an: Springer: 2011. p. 262–77.Google Scholar
 27.Laud P, Randmets J. A domainspecific language for lowlevel secure multiparty computation protocols In: Ray I, Li N, Kruegel C, editors. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security, Denver, CO, USA, October 126, 2015. New York: ACM: 2015. p. 1492–503. https://doi.org/10.1145/2810103.2813664.Google Scholar
 28.Bogdanov D, Laur S, Talviste R. A practical analysis of oblivious sorting algorithms for secure multiparty computation, vol. 8788 In: Bernsmed K, FischerHübner S, editors. Secure IT Systems  19th Nordic Conference, NordSec 2014, Tromsø, Norway, October 1517, 2014, Proceedings. Lecture Notes in Computer Science. Cham: Springer: 2014. p. 59–74. https://doi.org/10.1007/9783319115993.Google Scholar
 29.NTL. A Library for doing Number Theory. http://www.shoup.net/ntl/. Accessed 7 July 2018.
 30.Laud P. A private lookup protocol with low online complexity for secure multiparty computation, vol. 8958 In: Hui LCK, Qing SH, Shi E, Yiu SM, editors. Information and Communications Security  16th International Conference, ICICS 2014, Revised Selected Papers. Lecture Notes in Computer Science. Cham: Springer: 2014. p. 143–57. https://doi.org/10.1007/9783319219660.Google Scholar
 31.Bogdanov D, Kamm L, Kubo B, Rebane R, Sokk V, Talviste R. Students and taxes: a privacypreserving study using secure computation. PoPETs. 2016; 2016(3):117–35.Google Scholar
 32.Damgård I, Pastro V, Smart NP, Zakarias S. Multiparty computation from somewhat homomorphic encryption, vol. 7417 In: SafaviNaini R, Canetti R, editors. CRYPTO. Lecture Notes in Computer Science. Berlin: Springer: 2012. p. 643–62.Google Scholar
 33.Keller M, Orsini E, Scholl P. Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, Vienna, Austria, October 2428, 2016 In: Weippl ER, Katzenbeisser S, Kruegel C, Myers AC, Halevi S, editors. New York: ACM: 2016. p. 830–42. https://doi.org/10.1145/2976749.2978357.
Copyright information
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.