Detecting Backdoors in Pre-trained Encoders