Insufficient training data is a serious problem in all domains related to bioinformatics. Transfer learning is a promising tool to solve this problem, which relaxes the hypothesis that training data must be independent and identically distributed with the test data. We construct a sophisticated electroencephalography (EEG) signal representation and obtain an efficient EEG feature extractor through manifold constraints-based joint adversarial training with training data from other domains. EEG signal is more easily distinguished in the feature space mapped by the feature extractor. Negative transfer is one of the most challenging problems in transfer learning. In our approach, we apply manifold constraints to overcome this problem, which can avoid the geometric manifolds in the target domain being destroyed. The experiments demonstrate that our approach has many advantages when applied to EEG classification tasks.