Probabilistic frameworks for privacy-aware data mining

Often several cooperating parties would like to have a global view of their joint data for various data mining objectives, but cannot reveal the contents of individual records due to privacy, ownership or competitive considerations. In this talk, we present a probabilistic framework for resolving such seemingly contradictory goals. Rather than sharing parts of the original or perturbed data, the framework shares the parameters of suitable probabilistic models built at each local data site. We mathematically show that the best representative of all the data is a certain ldquomeanrdquo model, and empirically show that this model can be approximated quite well by generating artificial samples from the underlying distributions using Markov chain Monte Carlo techniques, and then fitting a combined global model with a chosen parametric form to these samples. We also propose a new measure that quantifies privacy in such situations based on information theoretic concepts, and show that decreasing privacy leads to a higher quality of the combined model and vice versa. The method can also be applied to situations where different local datasets may not have identical features by using certain maximum likelihood and maximum entropy principles. We provide empirical results on different data types with continuous vector, categorical and directional attributes to highlight the generality of our framework. The results show that high quality distributed clustering or classification can be achieved with little privacy loss and low communication cost.