Jian Zhang

Associate Professor

Division of Computer Science,
Louisiana State University

3272K Patrick Taylor Hall

Tel: (225)-578-8353

Email: zhang at csc dot lsu dot edu


Some Recent Research Projects:

  • Deep neural network for detection of Android malware

Malware targeting mobile devices is a pervasive problem in modern life. The detection of malware is essentially a software classification problem In this project, we investigated the effectiveness of Deep Neural Networks (DNNs) for classification of Android applications. A random walk was conducted on the code to generate sequences of event that are intrinsic to the program. Leveraging the ability of DNNs to learn complex and flexible features, we designed a Convolutional Neural Network (CNN) to detect malware based on the event sequence. We tested and compared our CNN to a recurrent neural network (LSTM) and other n-gram based methods. Both CNN and LSTM significantly outperformed n-gram based methods. Surprisingly, the performance of our CNN is also much better than that of the LSTM, which is considered a natural choice for sequential data.
code graph

Code graph on which we conduct random walk

dnn structure

Structure of the sequence CNN

  • Prediction of user attributes from tweets

User attributes such as age, and education are valuable information for various online services such as personalized recommendation, marketing, public health and social study. We consider the problem of inferring these attributes from a collection of tweets generated by online users. We designed a model that uses deep neural network to discover text patterns of different sizes and complexities and combines attention-aware DNN with max-margin classification. Experiment results on real-world datasets show that the model outperforms traditional text classification methods such as SVM and other deep neural-network models such as regular LSTM and CNN.
DNN Structure

DNN for prediction of user attributes