Privacy-Preserving Genetic Algorithms for Rule Discovery

Decision tree induction algorithms generally adopt a greedy approach to select attributes in order to optimize some criteria at each iteration of the tree induction process. When a decision tree has been constructed, a set of decision rules may be correspondingly derived. Univariate decision tree induction algorithms generally yield the same tree regardless of how many times it is induced from the same training data set. Genetic algorithms have been shown to discover a better set of rules, albeit at the expense of efficiency. In this paper, we propose a protocol for secure genetic algorithms for the following scenario: Two parties, each holding an arbitrarily partitioned data set, seek to perform genetic algorithms to discover a better set of rules without disclosing their own private data. The challenge for privacy-preserving genetic algorithms is to allow the two parties to securely and jointly evaluate the fitness value of each chromosome using each party's private data but without compromising their data privacy. We propose a new protocol to address this challenge that is correct and secure. The proposed protocol is not only privacy-preserving at each iteration of the genetic algorithm, the intermediate results generated at each iteration do not compromise the data privacy of the participating parties.