Device-free localization (DFL) technology estimates the target position by analyzing the signal fluctuations between wireless sensor nodes. Its core is establishing a position-signal mapping relationship using the characteristic signal changes caused by target movement. However, in practical applications, dynamic environmental changes cause the performance of the pre-built fingerprint database to deteriorate or even fail. This study proposes a robustness enhancement framework based on knowledge distillation (KD). It adopts a dual-input architecture of original data and noise-injected data. It constrains the output consistency under different noise conditions through Kullback-Leibler (KL) divergence to achieve cross-modal knowledge transfer from the clean data domain to the perturbed data domain. The experiment uses a multi-level noise pollution scenario to verify the performance of the model. The results show that under the conditions of −15 dB signal-to-noise ratio and 0.6 impulse noise ratio, this method maintains a positioning accuracy of 43.4% and 63.57%, respectively, compared with the traditional scheme, verifying the effectiveness of the knowledge transfer strategy. This framework provides new methodological support for DFL systems that can operate sustainably in dynamic environments, and significantly improves the environmental adaptability of the model through knowledge distillation.