In genome-wide interaction studies, to detect gene-gene interactions, most methods are divided into two folds: single nucleotide polymorphisms (SNP) based and gene-based methods. Basically, the methods based on the gene are more effective than the methods based on a single SNP. Recent years, while the kernel canonical correlation analysis (Classical kernel CCA) based U statistic (KCCU) has proposed to detect the nonlinear relationship between genes. To estimate the variance in KCCU, they have used resampling based methods which are highly computationally intensive. In addition, classical kernel CCA is not robust to contaminated data. We, therefore, first discuss robust kernel mean element, the robust kernel covariance, and cross-covariance operators. Second, we propose a method based on influence function to estimate the variance of the KCCU. Third, we propose a nonparametric robust KCCU method based on robust kernel CCA, which is designed for contaminated data and less sensitive to noise than classical kernel CCA. Finally, we investigate the proposed methods to synthesized data and imaging genetic data set. Based on gene ontology and pathway analysis, the synthesized and genetics analysis demonstrate that the proposed robust method shows the superior performance of the state-of-the-art methods.