Bad data outliers and malicious corruption in Phasor Measurement Unit (PMU) data having signature similar to that of a highly nonlinear event-induced oulier can challenge reliable event detection when linear principal component analysis (PCA)-based metrics are used. This paper presents a moving window based kernel PCA approach for accurately detecting event-induced outliers in presence of such corruptions in data. It is demonstrated that with appropriate tuning of kernel parameters, the change in the square of the norm of principal component score between successive windows along the direction of maximum variance in feature space can be used as a metric for corruption-resilient detection of event-induced outliers. Analytical justification for the same is provided along with a bound on this change. The performance of the proposed metric is validated on both synthetic data and field measurements.