Interactions among multiple genetic variants are likely to affect risk for human complex disease. It is increasingly recognized that the identification of interactions will not only increase the power to detect disease-associated variants, but will also help elucidate biological pathways that underlie diseases. In this article, we propose a two-stage method for detecting gene-gene interactions. In the first stage, using a model selection method, that is, support vector machines (SVM) with L1 penalty, we identify the most promising single-nucleotide polymorphisms (SNPs) and interactions. In the second stage, we apply logistic regression and ensure a valid type I error by excluding non-significant candidates after Bonferroni correction. We analyze a published case-control dataset where our method successfully identified an interaction term which was not discovered in previous studies.