Authentic Publication The TRUTHSAYER Project Chip Martel Premkumar

Authentic Publication The TRUTHSAYER Project Chip Martel Premkumar Devanbu Michael Gertz April Kwong Glen Nuckolls Stuart Stubblebine Department of Computer Science, University of California, Davis http: //truthsayer. cs. ucdavis. edu

Databases Play a Vital Role 1) Commerce: credit card data, find goods 2) Financial: Investment sites 3) Health: treatments, doctors/credentials, drugs 4) Many more

Answering queries Server Integrity? Correct Query processing? Performance? Reliability? Data Query User Answers Database

Goals • Correct and complete answers (with assurance) • Efficient Protocols

Example Queries • Is Credit card number 5543… Valid? • List all Hong Kong to San Francisco flights. • Find Digital cameras with 3 -5 Mega-pixels, and cost < $200 • List all bars within one mile of HKU

What is a Correct Answer? • We assume a trusted Data Owner with the official copy of the Database: Defines the “correct answer”

What is a Correct Answer? • We assume a trusted Data Owner with the official copy of the Database: Defines the “correct answer” • Problems with a single Data Owner: 1) May not want/be able to answer queries 2) Hard to keep online DB secure 3) Scalability

Solution: Third-Party Servers • Third party sites (Publishers) get information from the Data Owner and answer queries • Example: Travel sites (Expedia, Travelocity, Orbitz) answer using government airline Data (FAA)

Server Replication Can I Trust This Server? Travelocity FAA Data Orbitz Expedia

Trust Issues • Sites have left out cheaper flights from non -preferred airlines (deliberate) • Sites may be corrupted: outside hacker or insider • Errors

Authentic Publication: The TRUTHSAYER project. Initially: for RDB (DBSEC 2000, Jnl. Comp. Sec. ) General Model for a Variety of Data (Algorithmica, 2004) Owner Data + Digest of Data Query Answer + Verification Object Publisher

Talk Outline Introduction Background--- Merkle Trees Range Queries (Multi-attribute Queries) A General Model for Authenticated Data Structures • Conclusion • •

Authentic Publication 1) A trusted Owner digests the Data Set, and signs it. 2) Untrusted Publishers receive the data & signature. 3) Clients submit queries to untrusted Publishers. 4) Publishers return Answers (A), and Verification Objects (A+ VO) 5) Clients use A + VO to Prove the answer is correct/complete. Protocol is correct, and secure.

Verifying answers • • Protocol provides: Correctness: Returns exact elements matching the query. Completeness: Returns all elements matching query. Security: Cheating is infeasible. Efficiency: Overhead is low. Recall: No signatures!!

Merkle hashing a data set. h(d 1) • Leaves: data in some lexical order. • One way hash function h; h 1= h(d 1) • Bottom-up hashing, starting with data • Root hash value = the digest of the data set.

Merkle Trees • Classic use: prove that data value d is in the data set • Solves: Is Credit card number 5543… Valid? • But also can verify all items in a range: e. g. camcorders from $400 to $900

Verifying a Range 1 3 5 6 8 10 11 15 q To Show that q =(5, 6, 8) is the Answer to 4<d <10: Used Lower Bound 3, Upper Bound 10 and starred hash values to compute/verify root hash.

Verifying a Range 1 3 5 6 8 10 11 q Query: 4<d <10: Answer: 5, 6, 8 (in practice, key + data) Verification Object: [( (h(1), 3), (5, 6) ) ( (8, 10), *) ] 15

Authentic Publication Hash Digest Merkle Tree

Security Property • If the Answer and VO are correct, user accepts

Security Property • User accepts an Invalid answer only if a specific collision in h is found (provable): h(x, y)= z in a correct VO (x, y, z are the hash values of tree nodes), VO uses different x’, y’ with h(x’, y’)=z

Good Features • Proofs are short (size proportional to tree height and answer size). • Use hashes, a fast cryptographic operation • Proofs as easy to compute as finding the answer • No secret keys: hash function and digests all are public (no insider attack once data set is digested).

Extensions • Want to handle more complex queries • Find Digital cameras with 3 -5 Mega pixels, and cost < $200 • List all bars within one mile of HKU

Multi-Attribute Queries • Model as a 2 -D Range query • Find points (x, y) with a<x<b wc<y<d (a, d) (b, d) Pixels (a, c) Cost (b, c)

2 -Dimensional range tree • Leaves are 2 D points, or 2 attributes (cost, pixels). Sorted by x-value in X-tree • A Y-tree for each internal node

Searching a 2 D-range Tree • Find (x, y) with 4 < x <50 AND 4 < y < 10 • All in Associated Y-trees Match x-range

Searching a 2 D-range Tree • Find pairs (x, y) with 4 < x <50 AND 4 < y < 10 • In X-tree: subtrees rooted at 5 and 13 • Search in Associated Y-trees

Searching a 2 D-range Tree • Find (x, y) with 4 < x <50 AND 4 < y < 10 • Answer: (12, 5) and (23, 8) AND values in 5’s Y-tree

Digesting a 2 D-range Tree • Digest each Y-tree as Merkle tree • Each internal node in the X-tree gets the hash of three values: two children and associated Y-tree value

Range Trees • Let k be the number of answers (out of n) • Search: O(k+ log 2 n) time, nlogn space • improve to O(k+ logn) time with extra pointers (can still get a hash digest) • VO (proof) size also O(k+logn) • Extend to d-dimensions (d-attribute query). Search time: O(k+log(d-1) n), VO size: same.

Authenticated Data Structures • Problem: May want to use a variety of efficient data-structures: w B-trees (reduce disk access) w Suffix arrays (string queries) w Geometric data structures (items within one mile) w Many more

Authenticated Data Structures • Solution: General method to digest a data structure (produce a single summary hash value). • Efficient: Proof size and construction time = search time. • Secure: Similar security property: break only with a specific collision in h

Search DAGS • Our general setting is any data structure modeled by: w A labeled Directed Acyclic Graph (DAG) w A search process that visits DAG nodes and determines which neighboring nodes to visit next (based on labels of visited nodes) This Models a wide range of structures

A Search DAG • Search starts at the unique source node s of in-degree zero • Digesting starts from the sinks (here u, v ): hash the associated values s b c a u v

A Search DAG • D(u): Digest of u • Node u data : du • D(u)= h(du) • D(v)= h(dv) s b c a u v

A Search DAG • Other Digests use data and successors s b • D(c) = h(dc, D(v) ) • D(b)=h(db, D(v), D(c)) • D(s) is DAG Digest c a u v

Verification for Search DAG • Traditional Merkle Tree verification is Bottom up (hash path values to root) • We use top down verification to simulate a correct search • Owner provides search procedure P and root digest D(s)

Authentic Publication D(s), P DAG, P

Verification Object for DAG • VO: information so User can reproduce the search (and thus verify answers) • “Lines” of VO match steps of P: • Data of a node and successor hashes w ds, D(v 1), D(v 2) … (successors of s) w dv 1 , D(u 1), D(u 2), … (successors of v 1)

An Example Search • Starts at s, then visits b then v s b • VO: w ds, D(a), D(b), D(c) (line 1) D(s) = h(ds, D(a), D(b), D(c)) So know data ds is OK. c a u v

An Example Search • Starts at s, process ds and decide b is next • VO: w ds, D(a), D(b), D(c) [line 1] w db, D(v), D(c) [line 2] If D(b)=h(db, D(v), D(c)) (using D(b) from line 1) w Data db is correct s b c a u v

Verified Search • The verified computation proceeds until all nodes in the actual search are visited (the VO has one line for each node visited). • The correct answer is now returned by search procedure P.

Verified Search • The verified computation takes time proportional to the original search (visits the same nodes). • Security Proof: shows that a User accepts the wrong answer only if a specific collision in hash function h used (e. g. D(b)=h(d’b, D’(v), D’(c))

Updates • Typically Digests are updated with work similar to the data structure’s update time (e. g. length of the search paths to updated items) • If updates are frequent, overall scheme doesn’t work well (can use time-stamped digests)

Generalizations • Allowing multiple Owners: often want to query data collected from several owners. Can be done, but now need to trust owners and data collector. • Privacy: VO’s may reveal information about the data set. Methods to conceal extra data.

Generalizations • I/O efficient digests/VO’s: can use a multiway tree to store multiple values in one disk block (still logically a binary tree for VO purposes, but stored more efficiently). • Top-down search DAG approach may be improved for specific data-structures (e. g. 2 D range trees)

Generalizations • Collections of structured data: XML documents (can answer path queries) • Relational operations (Joins, Selection, Projection) • Fancier Crypto operations (to reduce VO size)

References P. Devanbu, M. Gertz, C. Martel, and S. G. Stubblebine. Authentic Third Party Data Publication, 14 th IFIP 11. 3 Working Conf. in DB Security (DBSec 2000), Original Authentic Publication Paper A General Model for Authenticated Data Structures, Algorithmica, 2004 Many Data Structures and Search DAG ( above group and G. Nuckolls)

References Certifying Data from Multiple Sources, Proceedings of the 17 th Database Security Conference, 2003 Shows how to use multiple Owners Flexible authentication of XML documents, Journal Computer Security, 2004

Survey Chapters Li, Hadjieleftheriou, Kollios, Reyzin Authenticated Index Structures for Outsourced Databases(Overview of area and efficiency issues) R. Sion: Towards Secure Data Outsourcing Michael Gertz and Sushil Jajodia (eds. ): "Handbook of Database Security: Applications and Trends", Springer, 2007, to appear. Both in:

A. Anagnostopoulos, M. Goodrich, R. Tamassia, Persistent Authenticated Dictionaries and Their Applications (allows queries of prior DB versions) Authenticated Data Structures for Graph and Geometric Searching (fancy geometric data structures)

Pointer for more information http: //truthsayer. cs. ucdavis. edu

Conclusion • A single signed Digest, can authenticate answers to many queries • Secure against hackers and insiders • Can handle a wide range of data structures • Efficient protocols: fast query processing and small VO’s

Future Work • Better Update Mechanisms • Integration of Database optimization methods • Actual implementation (partly done by others), and evaluation