Interest measures – make sure that sensitive facts, if they exist, will be deemed uninteresting by algorithms Extra data – example, a “phone book” that contains extra entries. Still useful if goal is to find phone given name, but access to complete phone book doesn’t allow determining facts about (for example) department sizes. Performance – maybe not an issue for small amounts of data, but on large data sets (terabyte); exponential performance is an issue (disk limited) Note that we don’t have the same problem faced by (for example) the GPS military/civilian accuracy encoding. There, the goal is to make information (position) known to all, but just more clearly for some. Here, the information to be made known, and the information to be kept hidden, are completely different. A better analogy would be getting position from communications satellites (e.g. measuring delay). Introducing a small random delay will wreak havoc with trying to determine position by this method, but will not alter the information communicated.
Here I try to summarize entire cryptographic approaches to PPDM. Basically, lots of work done on applying secure Multi-party based ideas for PPDM. Generally, it is applied to distributed data mining. In most of the work recent work, it is assumed That the adversaries are semi-honest (i.e. they follow the protocol correctly). Only recently (including Kantarcioglu and Kardes paper that will be presented in the workshop) malicious model is discussed. It turns out that all these Different solutions are consist of few common secure subprotocols such as dot product and summation.
Perturbation is a very important technique in PPDM. This technique is to distort the data, but still keep some properties of the data which will be used for later data mining phase. Here listed are some perturbation techniques. Additive based approach is first proposed by Agrawal and Srikant, now has many various. Single one step plus may not enough to protect the privacy, we have proposed a two step model in ICDM 06. Multiplicative based approach, e.g. orthogonal transformation, geometry property is to rotate the data. E.g Chen and Liu ICDM 05. This transformation has the property to keep the Euclidean distance between any pair of data points, so some data mining tools can be directly applied, K-Nearest Neighbor Classifier(KNN), Support Vector Machines(SVM) and so on. The later approach has evaluated the privacy preserving in more detail, so proposed a random projection to a lower space, Liu and Kargupta TKDE2006 and Liu and Kargupta PKDD'06. Condensation and decomposition (Wandand Zhang ICDM06) are using some properties of matrix. In the decomposition area, Wavelet transformation is new. All these approaches are still in progress. Data swapping is a different approach, which transforms the data set by switching a subset of attributes between selected pairs (Fienberg et al 2003.
Make the PPDM approaches more fit in the real life situation is the trend for today’s research. We conducted intensive experiments with real-world data set, and give a applicability study in DKE07 paper. Reconstruction of the original data distribution not work very well with real life data. Distribution is a hard problem. When the distribution of the original data set is not hard, the method may work; but if the distribution of the original data is hard, the method not work well. It depends on the distribution! So we suggest should not use distribution as a meddle step. In our another work, we have tailed the data mining tools to fit the PPDM domain. That is try to directly mapping the data mining functions according to the noise addition method. Believe this is a fruitful direction for PPDM
Data Mining, Security and Privacy Prof. Bhavani Thuraisingham Prof. Murat Kantarcioglu Ms Li Liu (PhD Student – completing December 2007) The University of Texas at Dallas August 24, 2008
Integrate the data, build warehouses and federations
Develop profiles of terrorists, activities/threats
Mine the data to extract patterns of potential terrorists and predict future activities and targets
Find the “needle in the haystack” - suspicious needles?
Data integrity is important
Techniques have to SCALE
Data Mining Needs for Counterterrorism: Real-time Data Mining
Nature of data
Data arriving from sensors and other devices
Continuous data streams
Breaking news, video releases, satellite images
Some critical data may also reside in caches
Rapidly sift through the data and discard unwanted data for later use and analysis (non-real-time data mining)
Data mining techniques need to meet timing constraints
Quality of service (QoS) tradeoffs among timeliness, precision and accuracy
Presentation of results, visualization, real-time alerts and triggers
Data Mining for Real-time Threats Integrate data sources in real - time Build real - time models Examine Results in Real - time Report final results Data sources with information about terrorists and terrorist activities Mine the data Rapidly sift through data and discard irrelevant data
Only make a sample of data available so that an adversary is unable to come up with useful rules and predictive functions
Introduce random values into the data and/or results
Challenge is to introduce random values without significantly affecting the data mining results
Give range of values for results instead of exact values
Secure Multi-party Computation
Each party knows its own inputs; encryption techniques used to compute final results
Rules, predictive functions
Approach: Only make a sample of data available
Limits ability to learn good classifier
Cryptographic Approaches for Privacy Preserving Data Mining
Secure Multi-part Computation (SMC) for PPDM
Mainly used for distributed data mining.; Provably secure under some assumptions.; Learned models are accurate; Mainly semi-honest assumption (i.e. parties follow the protocols); Malicious model is also explored recently. (e.g. Kantarcioglu and Kardes paper in this workshop); Many SMC based PPDM algorithms share common sub-protocols (e.g. dot product, summation, etc. )
Still not efficient enough for very large datasets. (e.g. petabyte sized datasets ??); Semi-honest model may not be realistic;Malicious model is even slower
Possible new directions
New models that can trade-off better between efficiency and security; Game theoretic / incentive issues in PPDM; Combining anonymization and cryptographic techniques for PPDM
Perturbation Based Approaches for Privacy Preserving Data Mining
Goal: Distort data while still preserve some properties for data mining propose.
Goal: Achieve a high data mining accuracy with maximum privacy protection.
Our approach: Privacy is a personal choice, so should be
(Liu, Kantarcioglu and Thuraisingham ICDM’06)
Perturbation Based Approaches for Privacy Preserving Data Mining
The trend is to make PPDM approaches to reflect reality
We investigated perturbation based approaches with real-world data sets
We give a applicability study to the current approaches
Liu, Kantarcioglu and Thuraisingham, DKE 07
We found out,
The reconstruction of the original distribution may not work well with real-world data sets
Try to modify perturbation techniques, and adapt some data mining tools, e.g. Liu, Kantarcioglu and Thuraisingham, Novel decision tree – UTD technical report 06
Platform for Privacy Preferences (P3P): What is it?
P3P is an emerging industry standard that enables web sites t9o express their privacy practices in a standard format
The format of the policies can be automatically retrieved and understood by user agents
It is a product of W3C; World wide web consortium
When a user enters a web site, the privacy policies of the web site is conveyed to the user; If the privacy policies are different from user preferences, the user is notified; User can then decide how to proceed
Several major corporations are working on P3P standards including
Privacy for Assured Information Sharing Export Data/Policy Component Data/Policy for Agency A Data/Policy for Federation Export Data/Policy Component Data/Policy for Agency C Component Data/Policy for Agency B Export Data/Policy
Privacy Preserving Surveillance Raw video surveillance data Face Detection and Face Derecognizing system Suspicious Event Detection System Manual Inspection of video data Comprehensive security report listing suspicious events and people detected Suspicious people found Suspicious events found Report of security personnel Faces of trusted people derecognized to preserve privacy
Directions: Foundations of Privacy Preserving Data Mining
We proved in 1990 that the inference problem in general was unsolvable, therefore the suggestion was to explore the solvability aspects of the problem.
Can we do something similar for privacy?
Is the general privacy problem solvable?
What are the complicity classes?
What is the storage and time complicity
We need to explore the foundation of PPDM and related privacy solutions
Directions: Testbed Development and Application Scenarios
There are numerous PPDM related algorithms. How do they compare with each other? We need a testbed with realistic parameters to test the algorithms
It is time to develop real world scenarios where these algorithms can be utilized
Is it feasible to develop realistic commercial products or should each organization adapt product to suit their needs?
I as a user of Organization A send data about me to organization B, I read the privacy policies enforced by organization B
If I agree to the privacy policies of organization B, then I will send data about me to organization B
If I do not agree with the policies of organization B, then I can negotiate with organization B
Even if the web site states that it will not share private information with others, do I trust the web site
Note: while confidentiality is enforced by the organization, privacy is determined by the user. Therefore for confidentiality, the organization will determine whether a user can have the data. If so, then the organization van further determine whether the user can be trusted