Learning versus Refutation in Noninteractive Local Differential Privacy