DL-based detection has gradually become the most effective detection methods due to that there is a large amount of data available. Traditional signature-based methods are not robust and generalizable enough since their customized security mode. Yet, ML-based methods are heavily dependent on the hand-crafted engineering. DL-based detection methods can be simply classified into four categories, 1) FCN-based, 2) CNN-based, 3) RNN-based, and 4) GNN/GCN-based. 1) FCN-based method, especially MLP-based, consists of at least three dense layers, and can learn non-linear patterns of data. Its detection speed is fast, but lacking accuracy. 2) CNN-based method utilizes CNN architecture for characterizing apps followed with a fully connected layer as a classifier, which can learn useful context information. However, the code pattern is sequential, it cannot accurately learn the long-term sequence dependencies. 3) RNN-based method utilizes RNN, GRU or LSTM architecture for remembering previous knowledge, which can capture a larger and more complex code context and semantics. Due to the above advantages, it performs achieving high accuracy and robustness, but costing a lone time. 4) GNN/GCN-based method applies to graphs of apps for characterizing apps. The graphs can be in many forms such as control flow graph, data flow graph, function call graph, and so on. Thus, this method can capture structure information and achieves high performance. But, it is also liable to structural attack.