Differential Privacy on Linked Data: Theory and Implementation Yotam Aron Table of Contents • Introduction • Differential Privacy for Linked Data • SPIM implementation • Evaluation Contributions • Theory: how to apply differential privacy to linked data. • Implementation: privacy module for SPARQL queries. • Experimental evaluation: differential privacy on linked data. Introduction Overview: Privacy Risk • Statistical data can leak privacy. • Mosaic Theory: Different data sources harmful when combined. • Examples: • Netflix Prize Data set • GIC Medical Data set • AOL Data logs • Linked data has added ontologies and meta-data, making it even more vulnerable. Current Solutions • Accountability: • Privacy Ontologies • Privacy Policies and Laws • Problems: • Requires agreement among parties. • Does not actually prevent breaches, just a deterrent. Current Solutions (Cont’d) • Anonymization • Delete “private” data • K – anonymity (Strong Privacy Guarantee) • Problems • • • • Deletion provides no strong guarantees Must be carried out for every data set What data should be anonymized? High computational cost (k-anonimity is np-hard) Differential Privacy • Definition for relational databases (from PINQ paper): A randomized function K gives Ζ-differential privacy if for all data sets π·1 and π·2 differing on at most one record, and all π ⊆ π ππππ(πΎ), Pr[πΎ π·1 ∈ π] ≤ exp π × Pr[πΎ π·2 ∈ π] Differential Privacy • What does this mean? • Adversaries get roughly same results from π·1 and π·2 , meaning a single individual’s data will not greatly affect their knowledge acquired from each data set. How Is This Achieved? • Add noise to result. • Simplest: Add Laplace noise Laplace Noise Parameters • Mean = 0 (so don’t add bias) • Variance = βπ , π where βπ is defined, for a record j, as πππ₯π (|πΉ π· − πΉ π· − π |) • Theorem: For query Q result R, the output R + Laplace(0, is differentially private. βπ ) π Other Benefit of Laplace Noise • A set of queries each with sensitivity ππ will have an overall sensitivity of ππ • Implementation-wise, can allocate an “budget” Ζ for a client and for each query client specifies ππ < π to use. Benefits of Differential Privacy • • • • Strong Privacy Guarantee Mechanism-Based, so don’t have to mess with data. Independent of data set’s structure. Works well with for statistical analysis algorithms. Problems with Differential Privacy • Potentially poor performance • Complexity • Noise • Only works with statistical data (though this has fixes) • How to calculate sensitivity of arbitrary query without bruteforce? Theory: Differential Privacy for Linked Data Differential Privacy and Linked Data • Want same privacy guarantees for linked data without, but no “records.” • What should be “unit of difference”? • One triple • All URIs related to person’s URI • All links going out from person’s URI Differential Privacy and Linked Data • Want same privacy guarantees for linked data without, but no “records.” • What should be “unit of difference”? • One triple • All URIs related to person’s URI • All links going out from person’s URI Differential Privacy and Linked Data • Want same privacy guarantees for linked data without, but no “records.” • What should be “unit of difference”? • One triple • All URIs related to person’s URI • All links going out from person’s URI Differential Privacy and Linked Data • Want same privacy guarantees for linked data without, but no “records.” • What should be “unit of difference”? • One triple • All URIs related to person’s URI • All links going out from person’s URI “Records” for Linked Data • Reduce links in graph to attributes • Idea: • Identify individual contributions from a single individual to total answer. • Find contribution that affects answer most. “Records” for Linked Data • Reduce links in graph to attributes, makes it a record. P1 Knows P2 Person Knows P1 P2 “Records” for Linked Data • Repeated attributes and null values allowed P1 Knows P2 Loves Knows P3 P4 Knows “Records” for Linked Data • Repeated attributes and null values allowed (not good RDBMS form but makes definitions easier) Person Knows Knows Loves P1 P2 Null P4 P3 P2 P4 Null Query Sensitivity in Practice • Need to find triples that “belong” to a person. • Idea: • Identify individual contributions from a single individual to total answer. • Find contribution that affects answer most. • Done using sorting and limiting functions in SPARQL Example S1 • COUNT of places visited S2 P1 MA State of Residence P2 Visited S3 Example S1 • COUNT of places visited S2 P1 MA State of Residence P2 Visited S3 Example S1 • COUNT of places visited S2 P1 MA State of Residence P2 Visited S3 Answer: Sensitivity of 2 Using SPARQL • Query: (COUNT(?s) as ?num_places_visited) WHERE{ ?p :visited ?s } Using SPARQL • Sensitivity Calculation Query (Ideally): SELECT ?p (COUNT(ABS(?s)) as ?num_places_visited) WHERE{ ?p :visited ?s; ?p foaf:name ?n } GROUP BY ?p ORDER BY ?num_places_visited LIMIT 1 In reality… • LIMIT, ORDER BY, GROUP BY doesn’t work together in 4store… • For now: Don’t use LIMIT and get top answers manually. • I.e. Simulate using these keywords in python • Will affect results, so better testing should be carried out in the future. • Would like to keep it on sparql-side ideally so there is less transmitted data (e.g. on large data sets) (Side rant) 4store limitations • • • • Many operations not supported in unison E.g. cannot always filter and use “order by” for some reason Severely limits the types of queries I could use to test. May be desirable to work with a different triplestore that is more up-to-date (ARQ). • Didn’t because wanted to keep code in python. • Also had already written all code for 4store Problems with this Approach • Need to identify “people” in graph. • Assume, for example, that URI with a foaf:name is a person and use its triples in privacy calculations. • Imposes some constraints on linked data format for this to work. • For future work, look if there’s a way to automatically identify private data, maybe by using ontologies. • Complexity is tied to speed of performing query over large data set. • Still not generalizable to all functions. …and on the Plus Side • Model for sensitivity calculation can be expanded to arbitrary statistical functions. • e.g. dot products, distance functions, variance, etc. • Relatively simple to implement using SPARQL 1.1 Implementation: Design of Privacy System SPARQL Privacy Insurance Module • i.e. SPIM • Use authentication, AIR, and differential privacy in one system. • Authentication to manage Ζ-budgets. • AIR to control flow of information and non-statistical data. • Differential privacy for statistics. • Goal: Provide a module that can integrate into SPARQL 1.1 endpoints and provide privacy. Design OpenID Authentication Differential Privacy Module HTTP Server SPIM Main Process Triplestore User Data AIR Reasoner Privacy Policies HTTP Server and Authentication HTTP Server OpenID Authentication • HTTP Server: Django server that handles http requests. • OpenID Authentication: Django module. SPIM Main Process • Controls flow of information. • First checks user’s budget, then uses AIR, then performs final differentially-private query. SPIM Main Process AIR Reasoner AIR Reasoner Privacy Policies • Performs access control by translating SPARQL queries to n3 and checking against policies. • Can potentially perform more complicated operations (e.g. check user credentials) Differential Privacy Protocol Client Differential Privacy Module Scenario: Client wishes to make standard SPARQL 1.1 statistical query. Client has Ζ “budget” of overall accuracy for all queries. SPARQL Endpoint Differential Privacy Protocol Client Query, Ζ>0 Differential Privacy Module Step 1: Query and epsilon value sent to the endpoint and intercepted by the enforcement module. SPARQL Endpoint Differential Privacy Protocol Client Differential Privacy Module Step 2: The sensitivity of the query is calculated using a re-written, related query. Sens Query SPARQL Endpoint Differential Privacy Protocol Client Step 3: Actual query sent. Differential Privacy Module Query SPARQL Endpoint Differential Privacy Protocol Client Result and Noise Differential Privacy Module Step 4: Result with Laplace noise sent over. SPARQL Endpoint Experimental Evaluation Evaluation • Three things to evaluate: • Correctness of operation • Correctness of differential privacy • Runtime • Used an anonymized clinical database as the test data and added fake names, social security numbers, and addresses. Correctness of Operation • Can the system do what we want? • Authentication provides access control • AIR restricts information and types of queries • Differential privacy gives strong privacy guarantees. • Can we do better? Use Case Used in Thesis • Clinical database data protection • HIPAA: Federal protection of private information fields, such as name and social security number, for patients. • 3 users • Alice: Works in CDC, needs unhindered access • Bob: Researcher that needs access to private fields (e.g. addresses) • Charlie: Amateur researcher to whom HIPAA should apply • Assumptions: • Django is secure enough to handle “clever attacks” • Users do not collude, so can allocate individual epsilon values. Use Case Solution Overview • What should happen: • Dynamically apply different AIR policies at runtime. • Give different epsilon-budgets. • How allocated: • Alice: No AIR Policy, no noise. • Bob: Give access to addresses but hide all other private information fields. • Epsilon budget: E1 • Charlie: Hide all private information fields in accordance with HIPAA • Epsilon budget: E2 Use Case Solution Overview • Alice: No AIR Policy • Bob: Give access to addresses but hide all other private information fields. • Epsilon budget: E1 • Charlie: Hide all private information fields in accordance with HIPAA • Epsilon budget: E2 Example: A Clinical Database HTTP Server OpenID Authentication • Client Accesses triplestore via HTTP server. • OpenID Authentication verifies user has access to data. Finds epsilon value, Example: A Clinical Database AIR Reasoner Privacy Policies • AIR reasoner checks incoming queries for HIPAA violations. • Privacy policies contain HIPAA rules. Example: A Clinical Database Differential Privacy Module • Differential Privacy applied to statistical queries. • Statistical result + noise returned to client. Correctness of Differential Privacy • Need to test how much noise is added. • Too much noise = poor results. • Too little noise = no guarantee. • Test: Run queries and look at sensitivity calculated vs. actual sensitivity. How to test sensitivity? • Ideally: • Test noise calculation is correct • Test that noise makes data still useful (e.g. by applying machine learning algorithms). • Fort his project, just tested former • Machine learning APIs not as prevalent for linked data. • What results to compare to? Test suite • 10 queries for each operation (COUNT, SUM, AVG, MIN, MAX) • 10 different WHERE CLAUSES • Test: • Sensitivity calculated from original query • Remove each personal URI using “MINUS” keyword and see which removal is most sensitive Example for Sens Test • Query: PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX foaf: <http://xmlns.com/foaf/0.1#> PREFIX mimic: <http://air.csail.mit.edu/spim_ontologies/mimicOntology#> SELECT (SUM(?o) as ?aggr) WHERE{ ?s foaf:name ?n. ?s mimic:event ?e. ?e mimic:m1 "Insulin". ?e mimic:v1 ?o. FILTER(isNumeric(?o)) } Example for Sens Test • Sensitivity query: PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX foaf: <http://xmlns.com/foaf/0.1#> PREFIX mimic: <http://air.csail.mit.edu/spim_ontologies/mimicOntology#> SELECT (SUM(?o) as ?aggr) WHERE{ ?s foaf:name ?n. ?s mimic:event ?e. ?e mimic:m1 "Insulin". ?e mimic:v1 ?o. FILTER(isNumeric(?o)) MINUS {?s foaf:name "%s"} } % (name) Results Query 6 - Error Runtime • Queries were also tested for runtime. • Bigger WHERE clauses • More keywords • Extra overhead of doing the calculations. Results Query 6 - Runtime Interpretation • Sensitivity calculation time on-par with query time • Might not be good for big data • Find ways to reduce sensitivity calculation time? • AVG does not do so well… • • • • Approximation yields too much noise vs. trying all possibilities Runs ~4x slower than simple querying Solution 1: Look at all data manually (large data transfer) Solution 2: Can we use NOISY_SUM / NOISY_COUNT instead? Conclusion Contributions • Theory on how to apply differential privacy to linked data. • Overall privacy module for SPARQL queries. • Limited but a good start • Experimental implementation of differential privacy. • Verification that it is applied correctly. • Other: • Updated sparql to n3 translation to Sparql version 1.1 • Expanded upon IARPA project to create policies against statistical queries. Shortcomings and Future Work • Triplestores need some structure for this to work • Personal information must be explicitly defined in triples. • Is there a way to automatically detect what triples would constitute private information? • Complexity • Lots of noise for sparse data. • Can divide data into disjoint sets to reduce noise like PINQ does • Use localized sensitivity measures? • Third party software problems • Would this work better using a different Triplestore implementation? Diff. Privacy and an Open Web • How applicable is this to an open web? • High sample numbers, but potentially high data variance. • Sensitivity calculation might take too long, need to approximate. • Can use disjoint subsets of the web to increase number of queries with Ι budgets. Demo • air.csail.mit.edu:8800/spim_module/ References • Differential Privacy Implementations: • “Privacy Integrated Queries (PINQ)” by Frank McSherry: http://research.microsoft.com/pubs/80218/sigmod115mcsherry.pdf • “Airavat: Security and Privacy for MapReduce” by Roy, Indrajit; Setty, Srinath T. V. ; Kilzer, Ann; Shmatikov, Vitaly; and Witchel, Emmet: http://www.cs.utexas.edu/~shmat/shmat_nsdi10.pdf • “Towards Statistical Queries over Distributed Private User Data” by Chen, Ruichuan; Reznichenko, Alexey; Francis, Paul; Gehrke, Johannes: https://www.usenix.org/conference/nsdi12/towardsstatistical-queries-over-distributed-private-user-data References • Theoretical Work • “Differential Privacy” by Cynthia Dwork: http://research.microsoft.com/pubs/64346/dwork.pdf • “Mechanism Design via Differential Privacy” by McSherry, Frank; and Talwar, Kunal: http://research.microsoft.com/pubs/65075/mdviadp.pdf • “Calibrating Noise to Sensitivity in Private Data Analysis” by Dwork, Cynthia; McSherry, Frank; Nissim, Kobbi; and Smith, Adam: http://people.csail.mit.edu/asmith/PS/sensitivity-tccfinal.pdf • “Differential Privacy for Clinical Trail Data: Preliminary Evaluations”, by Vu, Duy; and SlavkoviΔ, Aleksandra: http://sites.stat.psu.edu/~sesa/Research/Papers/padm09sesaSep 24.pdf References • Other • “Privacy Concerns of FOAF-Based Linked Data” by Nasirifard, Peyman; Hausenblas, Michael; and Decker, Stefan: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.153.5 772 • “The Mosaic Theory, National Security, and the Freedom of Information Act”, by David E. Pozen http://www.yalelawjournal.org/pdf/115-3/Pozen.pdf • “A Privacy Preference Ontology (PPO) for Linked Data”, by Sacco, Owen; and Passant, Alexandre: http://ceur-ws.org/Vol813/ldow2011-paper01.pdf • “k-Anonimity: A Model for Protecting Privacy”, by Latanya Sweeney: http://arbor.ee.ntu.edu.tw/archive/ppdm/Anonymity/SweeneyK A02.pdf References • Other • “Approximation Algorithms for k-Anonimity”, by Aggarwal, Gagan; Feder, Tomas; Kenthapadi, Krishnaram; Motwani, Rajeev; Panigraphy, Rina; Thomas, Dilys; and Zhu, An: http://research.microsoft.com/pubs/77537/k-anonymity-jopt.pdf Appendix: Results Q1, Q2 Q1 COUNT Error Q2 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.020976 0.05231 Query_Time Sens_Calc_Time 0 0.015823126 0.011798859 0 0.010298967 0.01198101 868.8379 0.010334969 0.04432416 0 0.010645866 0.012124062 0 0.010524988 0.012120962 Appendix: Results Q3, Q4 Q3 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.007927895 0.00800705 0 0.007529974 0.007997036 375.8253 0.00763011 0.030416012 0 0.007451057 0.008117914 0 0.007512093 0.008100986 Q4 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.01048708 0.012546062 0 0.01123786 0.012809038 860.91 0.011286974 0.048202038 0 0.01145792 0.01297307 0 0.011392117 0.012881041 Appendix: Results Q5, Q6 Q5 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.08081007 0.098078012 0 0.085678816 0.097680092 115099.5 0.087270975 0.373119116 0 0.084903955 0.097922087 0 0.083213806 0.098366022 Q6 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.136605978 0.153807878 0 0.139995098 0.155878067 115118.4 0.139881134 0.616436958 0 0.148360014 0.160467148 0 0.144635916 0.158998966 Appendix: Results Q7, Q8 Q7 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.006100178 0.004678965 0 0.004260063 0.004747868 0 0.004283905 0.017117977 0 0.004103184 0.004703999 0 0.004188061 0.004717112 Q8 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.002182961 0.002643108 0 0.002092123 0.002592087 0 0.002075911 0.002662182 0 0.00207901 0.002576113 0 0.002048969 0.002597094 Appendix: Results Q9, Q10 Q9 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.004920959 0.010298014 0 0.004822016 0.010312796 0.00037 0.004909992 0.024574041 0 0.004843235 0.01032114 0 0.004893064 0.010319948 Q10 COUNT SUM AVG MAX MIN Error Query_Time Sens_Calc_Time 0 0.012365818 0.014447212 0 0.013066053 0.014631987 860.91 0.013166904 0.056000948 0 0.013354063 0.014893055 0 0.013329029 0.014914989