Differentially Private Data Release for Data Mining Benjamin C.M. Fung Concordia University Montreal, QC, Canada Noman Mohammed Concordia University Montreal, QC, Canada Rui Chen Concordia University Montreal, QC, Canada Philip S. Yu University of Illinois at Chicago, IL, USA
2 Outline Overview Differential privacy Related Work Our Algorithm Experimental results Conclusion 2
3 Overview 3 Privacy model Anonymization algorithm Data utility
4 Contributions Proposed an anonymization algorithm that provides differential privacy guarantee G eneralization-based algorithm for differentially private data release Proposed algorithm can handle both categorical and numerical attributes Preserves information for classification analysis 4
5 Outline Overview Differential privacy Related Work Our Algorithm Experimental results Conclusion 5
6 Differential Privacy [DMNS06] 6 A non-interactive privacy mechanism A gives ε -differential privacy if for all neighbour D and D’, and for any possible sanitized database D* Pr A [A(D) = D*] ≤ exp(ε) × Pr A [A(D’) = D*] DD’ D and D’ are neighbors if they differ on at most one record
7 Laplace Mechanism [DMNS06] 7 For example, for a single counting query Q over a dataset D, returning Q(D) + Laplace(1/ε) maintains ε -differential privacy. ∆f = max D,D’ ||f(D) – f(D’)|| 1 For a counting query f: ∆f =1
8 Given a utility function u : ( D × T ) → R for a database instance D, the mechanism A, A(D, u) = return t with probability proportional to exp(ε×u(D, t)/2 ∆u) gives ε -differential privacy. Exponential Mechanism [MT07] 8
9 Composition properties 9 Sequential composition ∑ i ε i –differential privacy Parallel composition max( ε i )–differential privacy
10 Outline Overview Differential privacy Related Work Our Algorithm Experimental results Conclusion 10
11 Two Frameworks Interactive: Multiple questions asked/answered adaptively Anonymizer
12 Two Frameworks Interactive: Multiple questions asked/answered adaptively Anonymizer Non-interactive: Data is anonymized and released
13 Related Work 13 A. Blum, C. Dwork, F. McSherry, and K. Nissim. Practical privacy: The SuLQ framework. In PODS, A. Friedman and A. Schuster. Data mining with differential privacy. In SIGKDD, Is it possible to release data for classification analysis ?
14 Why Non-interactive framework ? 14 Disadvantages of interactive approach: Database can answer a limited number of queries Big problem if there are many data miners Provide less flexibility to perform data analysis
15 Non-interactive Framework 0 + Lap(1/ ε ) 15
16 For high-dimensional data, noise is too big 0 + Lap(1/ ε ) 16 Non-interactive Framework
17 Non-interactive Framework
18 Outline Overview Differential privacy Related Work Our Algorithm Experimental results Conclusion 18
19 JobAgeClassCount Any_Job[18-65)4Y4N8 Artist[18-65)2Y2N4 Professional[18-65)2Y2N4 Age [18-65) [18-40)[40-65) Artist[18-40)2Y2N4Artist[40-65)0Y0N0 Anonymization Algorithm [18-30)[30-40) 19 Professional[18-40)2Y1N3Professional[40-65)0Y1N1 Job Any_Job ProfessionalArtist EngineerLawyerDancerWriter
20 Candidate Selection we favor the specialization with maximum Score value First utility function: ∆u = Second utility function: ∆u = 1 20
21 Split Value The split value of a categorical attribute is determined according to the taxonomy tree of the attribute How to determine the split value for numerical attribute ? 21
22 Split Value The split value of a categorical attribute is determined according to the taxonomy tree of the attribute How to determine the split value for numerical attribute ? AgeClass 60 Y 30 N 25 Y 40 N 25 Y 40 N 45 N 25 Y
23 Anonymization Algorithm O(A pr x|D|log|D|) O(|candidates|) O(|D|) O(|D|log|D|) O(1) 23
24 Anonymization Algorithm O(A pr x|D|log|D|) O(|candidates|) O(|D|) O(|D|log|D|) O(1) O((A pr +h)x|D|log|D|) 24
25 Outline Overview Differential privacy Related Work Our Algorithm Experimental results Conclusion 25
26 Experimental Evaluation Adult: is a Census data (from UCI repository) 6 continuous attributes. 8 categorical attributes. 45,222 census records 26
27 Data Utility for Max 27
28 Data Utility for InfoGain 28
29 Comparison 29
30 Scalability 30
31 Outline Overview Differential privacy Related Work Our Algorithm Experimental results Conclusion 31
32 Differentially Private Data Release Generalization-based differentially private algorithm Provides better utility than existing techniques Conclusions 32
33 Q&A Thank You Very Much 33