Toward Stronger Textual Attack Detectors