Heterogeneous Defect Prediction (HDP) utilises data from other projects with heterogeneous metric sets to predict potentially defective modules in target project. Several existing HDP methods utilize feature engineering such as matching or transforming metrics to unify metrics and improve prediction performance. However, these methods do not perform well due to the large data distribution differences caused by heterogeneous metrics. We refer to this as the feature transferability issue. In response to this issue, this paper proposes a HDP method based on autoencoder and adversarial learning, abbreviated as EA-HDP. First, autoencoder is utilized to jointly encode the data of the source and target project, resulting in reconstructed feature vectors with implicit features. Then, using adversarial learning to reduce the implicit feature distribution difference between projects. Additionally, EA-HDP incorporates maximum mean discrepancy to further learn their common feature representation effectively. By jointly training the autoencoder, the adversarial network, and the classifie, EA-HDP can alleviate the feature transferability issue. Experiments demonstrate that EA-HDP outperforms baseline models.