Radio frequency (RF) fingerprint technology is utilized for wireless device identification, extensively employed in the internet of things (IoT). The operating environment for IoT devices is challenging, with pervasive noise and distortion on the signals which blur the feature space of RF fingerprints. Consequently, the model accuracy obtained through training at high signal-to-noise ratio (SNR) scenarios decreases with the low SNR of the received signals in testing. To solve the noise domain adaptation problem, an anti-noise scheme is proposed to enhance identification accuracy of RF fingerprint at varying SNRs. The squared cross power spectral density (SCPSD) features are first proposed to obtain the same RF fingerprint representation. Subsequently, the specific effect of noise on SCPSD is theoretically derived and the rationality of the scheme is demonstrated through simulation experiments. Finally, 60 off-the-shelf ZigBee devices are employed to evaluate the performance of the anti-noise algorithm. The experimental results show that employing the random subspace k-nearest neighbors (RSKNN) classifier not only effectively classifies devices with multi-cluster feature, but combined with the anti-noise scheme can significantly improve the accuracy by approximately 46% for SNRs not less than 5 dB.