Due to the significant individual variations in EEG, directly applying classifiers trained on known data from multiple subjects to unknown subjects can lead to performance degradation. For the specific task of adaptation of motor imagery (MI) decoding, we propose an improved adversarial DL model to facilitate the classification results of target subject by combining the use of an efficient backbone network and an advanced loss function. In particular, the approach simultaneously optimizes three essential components: a feature extractor, a domain discriminator, and a classifier. To confirm our model's efficacy, it was thoroughly assessed using the BCI Competition IV Datasets 2a and 2b. Results show that our model can effectively reduce feature discrepancies across different domains and improves the model's generalization capability.