Approaches to distributed privacy protecting data mining

Transcription

1 Approaches to distributed privacy protecting data mining Bartosz Przydatek CMU Approaches to distributed privacy protecting data mining p.1/11

2 Introduction Data Mining and Privacy Protection conflicting goals Approaches to distributed privacy protecting data mining p.2/11

3 Introduction Data Mining and Privacy Protection conflicting goals Conflict Resolution: Inference Control Inference Control Techniques Controlled Release Input/Output Perturbation Query Restriction & Auditing... Approaches to distributed privacy protecting data mining p.2/11

4 Introduction Data Mining and Privacy Protection conflicting goals Conflict Resolution: Inference Control Inference Control Techniques Controlled Release Input/Output Perturbation Query Restriction & Auditing... Distributed data mining old and new challenges Approaches to distributed privacy protecting data mining p.2/11

5 Outline What does privacy mean? Perturbation techniques Secure Multi-Party Computation (MPC) Privacy by secure MPC Conclusions Approaches to distributed privacy protecting data mining p.3/11

6 What Does Privacy Mean? Intuitively, it seems to be clear... Exact vs. partial disclosure Approaches to distributed privacy protecting data mining p.4/11

7 What Does Privacy Mean? Intuitively, it seems to be clear... Exact vs. partial disclosure Quantifying Privacy Interval width for a confidence level [Agrawal, Srikant 2000] Information theoretic approach [Agrawal, Aggarwal 2001] Game theoretic approach [Kleinberg, Papadimitriou, Raghavan 2001] Approaches to distributed privacy protecting data mining p.4/11

8 What Does Privacy Mean? Intuitively, it seems to be clear... Exact vs. partial disclosure Quantifying Privacy Interval width for a confidence level [Agrawal, Srikant 2000] Information theoretic approach [Agrawal, Aggarwal 2001] Game theoretic approach [Kleinberg, Papadimitriou, Raghavan 2001] Privacy in secure multi-party computation Approaches to distributed privacy protecting data mining p.4/11

9 Privacy by Perturbation Studied extensively in the context of single databases Can be applied in distributed setting Approaches to distributed privacy protecting data mining p.5/11

10 Privacy by Perturbation Studied extensively in the context of single databases Can be applied in distributed setting Various techniques randomized input distortion output perturbation k-anonymity [Sweeney 98] Approaches to distributed privacy protecting data mining p.5/11

11 Problems with Perturbations Bias, precision & consistency Can be computationally challenging Outlier removal & blurring the data detection of anomalies? Combining multiple versions of data released for different purposes Approaches to distributed privacy protecting data mining p.6/11

12 Secure Multi-Party Computation Introduced by Yao in 1982, inspired by coin-flipping (Blum) and mental poker (Shamir, Rivest, Adleman) m parties P 1,..., P m want to compute f(x 1,..., x m ), where x i is a private input of P i, without revealing more than necessary... Approaches to distributed privacy protecting data mining p.7/11

13 Secure Multi-Party Computation Introduced by Yao in 1982, inspired by coin-flipping (Blum) and mental poker (Shamir, Rivest, Adleman) m parties P 1,..., P m want to compute f(x 1,..., x m ), where x i is a private input of P i, without revealing more than necessary i.e., simulation of a trusted party! Approaches to distributed privacy protecting data mining p.7/11

14 Secure Multi-Party Computation Introduced by Yao in 1982, inspired by coin-flipping (Blum) and mental poker (Shamir, Rivest, Adleman) m parties P 1,..., P m want to compute f(x 1,..., x m ), where x i is a private input of P i, without revealing more than necessary i.e., simulation of a trusted party! A very general and powerful tool, various models Efficient completeness results: [Yao 86] (2-party), [GMW 87] (crypt.) and [BGW+CCD 88] (uncond.) Approaches to distributed privacy protecting data mining p.7/11

15 Privacy by Multi-Party Computation MPC creates a trusted party! Approaches to distributed privacy protecting data mining p.8/11

16 Privacy by Multi-Party Computation MPC creates a trusted party! Problems: Efficiency communication complexity Approaches to distributed privacy protecting data mining p.8/11

17 Privacy by Multi-Party Computation MPC creates a trusted party! Problems: Efficiency communication complexity Does it really solve the privacy problem? Approaches to distributed privacy protecting data mining p.8/11

18 Efficient MPC Solutions Efficient special purpose protocols Learning decision trees [Lindell, Pinkas 2000] Approaches to distributed privacy protecting data mining p.9/11

19 Efficient MPC Solutions Efficient special purpose protocols Learning decision trees [Lindell, Pinkas 2000] Private approximations Introduced by [FIMNSW 2000] A tradeoff between privacy and approximability [Halevi, Krauthgamer, Kushilevitz, Nissim, 2001] Some functions cannot be computed with low communication (set equality vs. set disjointness) Approaches to distributed privacy protecting data mining p.9/11

20 Efficient MPC Solutions Efficient special purpose protocols Learning decision trees [Lindell, Pinkas 2000] Private approximations Introduced by [FIMNSW 2000] A tradeoff between privacy and approximability [Halevi, Krauthgamer, Kushilevitz, Nissim, 2001] Some functions cannot be computed with low communication (set equality vs. set disjointness) A different approach to MPC? Approaches to distributed privacy protecting data mining p.9/11

21 Which queries preserve privacy? Approaches to distributed privacy protecting data mining p.10/11

22 Which queries preserve privacy? Query restriction query-set-size, query-set-overlap query auditing partitioning Approaches to distributed privacy protecting data mining p.10/11

23 Which queries preserve privacy? Query restriction query-set-size, query-set-overlap query auditing partitioning Query auditing efficient in simple cases a NP-hard problem in general [Kleinberg, Papadimitriou, Raghavan 2001] Approaches to distributed privacy protecting data mining p.10/11

24 Conclusions Privacy means... Various approaches, problem dependent Probably no the best single solution Still a lot of work to be done Approaches to distributed privacy protecting data mining p.11/11