Depends on the problem and the data. Accuracy could be used, but depending on the goals and costs of false positives vs false negatives something like precision or recall may be more appropriate, or using an ROC curve.