1. 1 Current Trends in Data Security Dan Suciu Joint work with Gerome Miklau

2. 2 Data Security Dorothy Denning, 1982: Data Security is the science and study of methods of protecting data (...) from unauthorized disclosure and modification Data Security = Confidentiality + Integrity

3. 3 Data Security Distinct from systems and network security –Assumes these are already secure Tools: –Cryptography, information theory, statistics, … Applications: –An enabling technology

4. 4 Outline Traditional data security Two attacks Data security research today Conclusions

5. 5 Traditional Data Security Security in SQL = Access control + Views Security in statistical databases = Theory

6. 6 Access Control in SQL GRANT privileges ON object TO users [WITH GRANT OPTIONS] privileges = SELECT | INSERT | DELETE |... object = table | attribute REVOKE privileges ON object FROM users [CASCADE ] [Griffith&Wade'76, Fagin'78]

7. 7 Views in SQL A SQL View = (almost) any SQL query Typically used as: GRANT SELECT ON pmpStudents TO DavidRispoli CREATE VIEW pmpStudents AS SELECT * FROM Students WHERE…

8. 8 Summary of SQL Security Limitations: No row level access control Table creator owns the data: that’s unfair ! … or spectacular failure: Only 30% assign privileges to users/roles –And then to protect entire tables, not columns Access control = great success story of the DB community...

9. 9 Summary (cont) Most policies in middleware: slow, error prone: –SAP has 10**4 tables –GTE over 10**5 attributes –A brokerage house has 80,000 applications –A US government entity thinks that it has 350K Today the database is not at the center of the policy administration universe [Rosenthal&Winslett’2004]

10. 10 Security in Statistical DBs Goal: Allow arbitrary aggregate SQL queries Hide confidential data SELECT count(*) FROM Patients WHERE age=42 and sex=‘M’ and diagnostic=‘schizophrenia’ SELECT count(*) FROM Patients WHERE age=42 and sex=‘M’ and diagnostic=‘schizophrenia’ OK SELECT name FROM Patient WHERE age=42 and sex=‘M’ and diagnostic=‘schizophrenia’ SELECT name FROM Patient WHERE age=42 and sex=‘M’ and diagnostic=‘schizophrenia’ Not OK [ Adam&Wortmann’89]

11. 11 Security in Statistical DBs What has been tried: Query restriction –Query-size control, query-set overlap control, query monitoring –None is practical Data perturbation –Most popular: cell combination, cell suppression –Other methods, for continuous attributes: may introduce bias Output perturbation –For continuous attributes only [ Adam&Wortmann’89]

12. 12 Summary on Security in Statistical DB Original goal seems impossible to achieve Cell combination/suppression are popular, but do not allow arbitrary queries

13. 13 Outline Traditional data security Two attacks Data security research today Conclusions

14. 14 Search claims by: SQL Injection Your health insurance company lets you see the claims online: Now search through the claims : Dr. Lee First login: User: Password: fred ******** SELECT…FROM…WHERE doctor=‘Dr. Lee’ and patientID=‘fred’ [Chris Anley, Advanced SQL Injection In SQL]

15. 15 SQL Injection Now try this: Search claims by: Dr. Lee’ OR patientID = ‘suciu’; -- Better: Search claims by: Dr. Lee’ OR 1 = 1; -- …..WHERE doctor=‘Dr. Lee’ OR patientID=‘suciu’; --’ and patientID=‘fred’

16. 16 SQL Injection When you’re done, do this: Search claims by: Dr. Lee’; DROP TABLE Patients; --

17. 17 SQL Injection The DBMS works perfectly. So why is SQL injection possible so often ? Quick answer: –Poor programming: use stored procedures ! Deeper answer: –Move policy implementation from apps to DB

18. 18 Latanya Sweeney’s Finding In Massachusetts, the Group Insurance Commission (GIC) is responsible for purchasing health insurance for state employees GIC has to publish the data: GIC(zip, dob, sex, diagnosis, procedure,...)

19. 19 Latanya Sweeney’s Finding Sweeney paid $20 and bought the voter registration list for Cambridge Massachusetts: GIC(zip, dob, sex, diagnosis, procedure,...) VOTER(name, party,..., zip, dob, sex) GIC(zip, dob, sex, diagnosis, procedure,...) VOTER(name, party,..., zip, dob, sex)

20. 20 Latanya Sweeney’s Finding William Weld (former governor) lives in Cambridge, hence is in VOTER 6 people in VOTER share his dob only 3 of them were man (same sex) Weld was the only one in that zip Sweeney learned Weld’s medical records ! zip, dob, sex

21. 21 Latanya Sweeney’s Finding All systems worked as specified, yet an important data has leaked How do we protect against that ? Some of today’s research in data security address breaches that happen even if all systems work correctly

22. 22 Summary on Attacks SQL injection: A correctness problem: –Security policy implemented poorly in the application Sweeney’s finding: Beyond correctness: –Leakage occurred when all systems work as specified

23. 23 Outline Traditional data security Two attacks Data security research today Conclusions

24. 24 Research Topics in Data Security Rest of the talk: Information Leakage Privacy Fine-grained access control Data encryption Secure shared computation

25. 25 FirstLastAgeRace HarryStone34Afr-Am JohnReyser36Cauc BeatriceStone47Afr-am JohnRamos22Hisp FirstLastAgeRace *Stone30-50Afr-Am JohnR*20-40* *Stone30-50Afr-am JohnR*20-40* Information Leakage: k-Anonymity Definition: each tuple is equal to at least k-1 others Anonymizing: through suppression and generalization Hard: NP-complete for supression only Approximations exists [Samarati&Sweeney’98, Meyerson&Williams’04]

26. 26 Information Leakage: Query-view Security Secret QueryView(s)Disclosure ? S(name)V(name,phone) S(name,phone) V1(name,dept) V2(dept,phone) S(name)V(dept) S(name) where dept=‘HR’ V(name) where dept=‘RD’ TABLE Employee(name, dept, phone) Have data: total big tiny none [Miklau&S’04, Miklau&Dalvi&S’05,Yang&Li’04]

27. 27 Summary on Information Disclosure The theoretical research: –Exciting new connections between databases and information theory, probability theory, cryptography The applications: –many years away [Abadi&Warinschi’05]

28. 28 Privacy “Is the right of individuals to determine for themselves when, how and to what extent information about them is communicated to others” More complex than confidentiality [Agrawal’03]

29. 29 Privacy Involves: Data Owner Requester Purpose Consent Example: Alice gives her email to a web service alice@a.b.com Privacy policy: P3P

30. 30 Hippocratic Databases DB support for implementing privacy policies. Purpose specification Consent Limited use Limited retention … [Agrawal’03, LeFevrey’04] alice@a.b.com Privacy policy: P3P Hippocratic DB Protection against: Sloppy organizations  Malicious organizations

31. 31 Privacy for Paranoids Idea: rely on trusted agents alice@a.b.com Agent aly1@agenthost.com lice27@agenthost.com foreign keys ? [Aggarwal’04] Protection against: Sloppy organizations Malicious attackers

32. 32 Summary on Privacy Major concern in industry –Legislation –Consumer demand Challenge: –How to enforce an organization’s stated policies

33. 33 Fine-grained Access Control Control access at the tuple level. Policy specification languages Implementation

34. 34 Policy Specification Language CREATE AUTHORIZATION VIEW PatientsForDoctors AS SELECT Patient.* FROM Patient, Doctor WHERE Patient.doctorID = Doctor.ID and Doctor.login = %currentUser CREATE AUTHORIZATION VIEW PatientsForDoctors AS SELECT Patient.* FROM Patient, Doctor WHERE Patient.doctorID = Doctor.ID and Doctor.login = %currentUser Context parameters No standard, but usually based on parameterized views.

35. 35 Implementation SELECT Patient.name, Patient.age FROM Patient WHERE Patient.disease = ‘flu’ SELECT Patient.name, Patient.age FROM Patient WHERE Patient.disease = ‘flu’ SELECT Patient.name, Patient.age FROM Patient, Doctor WHERE Patient.disease = ‘flu’ and Patient.doctorID = Doctor.ID and Patient.login = %currentUser SELECT Patient.name, Patient.age FROM Patient, Doctor WHERE Patient.disease = ‘flu’ and Patient.doctorID = Doctor.ID and Patient.login = %currentUser e.g. Oracle

36. 36 Two Semantics The Truman Model = filter semantics –transform reality –ACCEPT all queries –REWRITE queries –Sometimes misleading results The non-Truman model = deny semantics –reject queries –ACCEPT or REJECT queries –Execute query UNCHANGED –May define multiple security views for a user [Rizvi’04] SELECT count(*) FROM Patients WHERE disease=‘flu’

37. 37 Summary of Fine Grained Access Control Trend in industry: label-based security Killer app: application hosting –Independent franchises share a single table at headquarters (e.g., Holiday Inn) –Application runs under requester’s label, cannot see other labels –Headquarters runs Read queries over them Oracle’s Virtual Private Database [Rosenthal&Winslett’2004]

38. 38 Data Encryption for Publishing Users and their keys: Complex Policies: All authorized users: K user Patient: K pat Doctor: K dr Nurse: K nu Administrator : K admin All authorized users: K user Patient: K pat Doctor: K dr Nurse: K nu Administrator : K admin What is the encryption granularity ? Doctor researchers may access trials Nurses may access diagnostic Etc… Doctor researchers may access trials Nurses may access diagnostic Etc… Scientist wants to publish medical research data on the Web

39. 39 Data Encryption for Publishing An XML tree protection: JoeDoe 28 Seattle flu K user K pat  (K nu  K adm )K nu  K dr K dr K pat K master TylenolCandy [Miklau&S.’03] Doctor: K user, K dr Nurse: K user, K nu Nurse+admin: K user, K nu, K adm

40. 40 Summary on Data Encryption Industry: –Supported by all vendors: Oracle, DB2, SQL-Server –Efficiency issues still largely unresolved Research: –Hard theoretical security analysis [Abadi&Warinschi’05]

41. 41 Secure Shared Processing Alice has a database DB A Bob has a database DB B How can they compute Q(DB A, DB B ), without revealing their data ? Long history in cryptography Some database queries are easier than general case

42. 42 Secure Shared Processing [Agrawal’03] Alice Bob a b c dc d e h(a) h(b) h(c) h(d)h(c) h(d) h(e) Compute one-way hash Exchange h(c) h(d) h(e)h(a) h(b) h(c) h(d) What’s wrong ? Task: find intersection without revealing the rest

43. 43 Secure Shared Processing AliceBob a b c d c d e E B (c) E B (d) E B (e)E A (a) E A (b) E A (c) E A (d) commutative encryption: h(x) = E A (E B (x)) = E B (E A (x)) E A (a) E A (b) E A (c) E A (d) E B (c) E B (d) E B (e) EAEA EBEB h(c) h(d) h(e)h(a) h(b) h(c) h(d) EAEA EBEB h(c) h(d) h(e) [Agrawal’03]

44. 44 Summary on Secure Shared Processing Secure intersection, joins, data mining But are there other examples ?

45. 45 Outline Traditional data security Two attacks Data security research today Conclusions

46. 46 Conclusions Traditional data security confined to one server –Security in SQL –Security in statistical databases Attacks possible due to: –Poor implementation of security policies: SQL injection –Unintended information leakage in published data

47. 47 Conclusions State of the industry: –Data security policies: scattered throughout applications –Database no longer center of the security universe –Needed: automatic means to translate complex policies into physical implementations State of research: data security in global data sharing –Information leakage, privacy, secure computations, etc. –Database research community has an increased appetite for cryptographic techniques

48. 48 Questions ?