<p>Inertial gait recognition is attractive for its noninvasiveness and privacy preservation. However, real-world applications suffer from domain shifts induced by common covariates (clothing and shoe changes), which is also known as disguised inertial gait recognition. These covariates alter inertial gait dynamics in different ways: footwear changes mainly affect short-term foot-ground impact and step rhythm, whereas clothing changes may influence body swing and in-pocket smartphone motion over longer temporal intervals. In addition, conventional single-scale networks cannot capture gait dynamics at multiple temporal resolutions. To address these limitations, we propose a multi-scale local–global temporal network with domain adaptation (MS-LGNet) to improve inertial gait recognition under multiple covariate variations. First, a multi-scale local temporal extractor (MLTE) captures local features at three time scales, thereby capturing covariate-induced gait variations from fine to coarse temporal resolutions. Next, a global temporal dynamic modeling (GTDM) module aggregates the multi-scale features and models long-range temporal dependencies. Finally, feature-distribution alignment is enforced via an adversarial training mechanism, to mitigate covariate-induced domain shifts. On the DIGRD disguised inertial gait recognition dataset, MS-LGNet increases the classification accuracy by 5.83% and 5.98% under the clothing and shoe-type covariates, respectively. Single-covariate controlled experiments show that MS-LGNet achieves the highest scores in all settings. We further validate the method’s effectiveness on the public WhuGait dataset.</p>