MSc(CompSc) Dissertation Public Seminar - November 2019

Title:  Medial Axis-based Convolutional Neural Networks for 3D Shape Segmentation

Speaker:  Chan Wing Hei


There has been a growing interest of Graph Convolution Neural Networks (Graph CNNs) on 3D shape analysis due to its capability and effectiveness of feature extraction from irregular 3D shape representations such as point clouds and meshes. However, most of the existing shape analysis approaches are designed to operate on complete 3D shapes, which is not considered to be the most efficient, considering the large data volumes of complete and unprocessed 3D shapes. Meanwhile, Medial Axis Transform (MAT) provides a simplified shape descriptor of 3D geometric shapes, while preserving shape topology. Both MAT and Graph CNN are often adopted in various 3D shape analysis tasks by their own, but rarely together. This study proposes a novel Medial Axis-based 3D Shape Segmentation approach using Graph CNN, which is the first deep learning-based 3D shape segmentation approach using MAT of 3D shapes. We demonstrate that MAT provides useful clues for segmentation which are well-captured by our Graph CNN model. Also, we show that our model reduces input data size without sacrificing model performance.

About the Speaker:

Chan Wing Hei is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. His supervisor is Dr. Loretta Choi.

Title:  Wifi Password Cracking Based on PassGAN

Speaker:  Chen Yiming


According to the investigation, numerous WIFI password are too simple.  Because many people consider that wireless network are less important than other private property.  In fact, ignorance of wireless password strength can lead to wireless Network compromised, which give the attacker’s opportunity to monitor the wireless network traffic and steal the valuable information from the WIFI user.

Among the existing WIFI penetration testing tools, dictionary attack is always used to crack WIFI password.  However, a lot weak passwords are not included in the leaked password dataset but show the same features as the passwords in the leaked dataset.  The existing WIFI penetration tools can not find a way to predict these kinds of simple password.

In this dissertation, I introduce an AI password generation tools called PassGan and integrate it into one of the open source WIFI penetration test tools, Wifite.  Passgan learn how to generate the password from the leaked password dataset and can generate numerous similar passwords excluded from the dataset.  Wifite can make use of these generated password to test the WIFI password strength.  Furthermore, this dissertation also include a research on Passgan learning ability on different kinds of passwords.

About the Speaker:

Chen Yiming is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong.  His supervisor is Prof. SM Yiu.

Title:  An Improved 3D Model Version Control Framework for Efficient Rendering

Speaker:  Qian Ruiyuan


This dissertation proposes a new method to store and manage the 3D components in a 3D model VCS (version control system). The main contribution is the persistent octree, persistent N-ary tree, persistent block array and Fast Diff algorithm proposed for speeding up the basic operations in a VCS. Normally, Diff and Merge operations need O(N) time to perform. But in our proposed framework, Diff and Merge can be performed quite efficiently.

About the Speaker:

Qian Ruiyuan is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. His supervisor is Dr. Loretta Choi.

Title:  Study on Stock Networks

Speaker:  Song Yulu


The complex network has been applied in many fields since it was proposed. Portfolio optimization has been a hot topic in the financial field since 1950. In recent years, researchers try to construct stock complex network to solve the problems in portfolio optimization. The structure and the characters of the stock network are well studied but there isn’t a specific application of stock network.

In this research, I try to use the concept of network centrality to select stocks in the optimal portfolio. I build the stock network based on CSI 300 index data and S&P 500 index data. After building the stock network, I choose the most central nodes to form the portfolio. I try to combine centralities with different combination methods and choose the combination method with the best performance. The results show that this method can simplify portfolio optimization process and obtain better results than the previous stock network methods.

About the Speaker:

Song Yulu is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. Her supervisor is Dr. Zhiyi Huang.

Title:  Design and Development of HKU Short Video Mini Program

Speaker:  Wang Mingsheng


WeChat mini program was released in early 2017. Users can open applications without downloading. The development of mini program developer community is growing rapidly, many institutions and universities also have their own small programs.

Short videos can transmit more information that is difficult to be transmitted by text or images. Because numerous students and parents can not get access to YouTube or other Google software, some of them are interested in HKU but have a lack of channels to understand the life and scenery of HKU. This mini program is well suited to them.

This project is a full-stack product with two front-end (mini program and user admin) and two back-end (Spring MVC and Spring Boot). All of them will be able to be visited remotely.

In order to optimize the performance of the background, this project used a distributed backend structure instead of simple monolithic application. Make use of zookeeper for file operation monitoring. After testing, performance can be optimized about 40%. In addition, in the way of integrating the Spring framework and FFmpeg, we have studied how to optimize at the application level compared to the traditional direct approach, with about 30% optimization after testing.

About the Speaker:

Wang Mingsheng is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. His supervisor is Dr. TW Chim.

Title:  Zero-day Malwares Detection Based on Deep Learning

Speaker:  Wen Qiaokun


In recent years, scientists have been working on zero-day malware detection all the times. With the advancement of technology, malwares can always avoid the detection of anti-virus tools by means of packing, confusion, splitting, anti-sandboxing, etc. In most situations, different companies or institutes detect whether the files go through their systems are malwares by capturing packets. Because the limit length of packages, packets capture tools often fail to get the entire file. And also, the self-destruction mechanism of many malicious codes also makes us always get a fragment of a PE file only. This makes it impossible to use too many domain knowledges to get features of malware. Therefore, this paper proposes to detect malwares according to very small binary fragments of PE files by using a CNN-based model. In the end, for a random piece of continuous malicious code, we achieved an accuracy of up to 0.86 when the length of continuous fragments is 60,000 bytes. For non-contiguous and unordered random pieces of malicious code, we get an accuracy of up to 0.83 using only 1024 bytes(1KB) length fragments. And when using 60,000 bytes length fragment as the baseline, we can finally receive up to 0.91 accuracy.

About the Speaker:

Wen Qiaokun is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. Her supervisor is Dr. KP Chow.

Title:  Statistical Arbitrage Using Machine Learning

Speaker:  Xiong Lingchu


This research paper analyzes the performance of four machine learning techniques including Logistic Regression (LR), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Long Short-Term Memory Networks (LSTM), applied to the Chinese stock market as represented by the CSI 500 constituents from January 2016 to June 2018. At first, prediction performance on the direction of stock prices is evaluated. Two models for each Machine Learning technique are prepared, one has binary output labeled decrease and increase. The other has three-dimension output labeled decrease, increase and flat. After adding the new attribute “flat”, four ML models are apparently significantly, especially in their predictabilities and probabilities. Based on the results of forecasting, a statistical arbitrage strategy is further constructed. The back-test results are compared across four methods. The empirical results indicate Random Forest is the strongest technique, as it can beat the market and gain profits whether transaction costs are considered or not. Gradient Boosting Decision Tree could also be profitable prior to costs. However, once transaction costs are taken into account, only Random Forest could make money. Logistic Regression underperforms other ML techniques due to its poor ability to capture non-linear patterns in the capital market.

About the Speaker:

Xiong Lingchu is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. Her supervisor is Dr. Jingrui Zhang.

Title:  Mortality Risk Score Prediction Based on the Clinical Data Using Deep Learning

Speaker:  Yang Lingyan


The mortality risk prediction is a valuable research filed. All-cause mortality risk prediction only based on diagnose is a brand-new attempt. In the study, risk score prediction problem is simplified into a simple yes-or-no answer. Under the usage of deep learning, the problem can be solved by binary classification models. Data processing and model selection become the key of the system. The basic idea of the data processing is that the survival status of the later period is determined based on the diagnostic records of the previous period. Different processed datasets come from different time lengths. By comparison of results from several common machine learning models on these datasets, the data processing method and classification model are determined. From the experiment, the ANN classification model has good performance for the dataset with 2-year observation time and 1-year prediction interval.

About the Speaker:

Yang Lingyan is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. Her supervisor is Dr. Hao Luo.

Title:  Exercise Recommendation App for Elderly Users

Speaker:  Yang Yuxiang


Parkinson’s disease brings a lot of pains to patients and their families. The core signs and symptoms of Parkinson’s disease are tremor, slowed movement (bradykinesia), rigid muscles, impaired posture and balance, loss of automatic movements, speech changes and Writing changes. Medical treatments such as drugs and surgery, are common and traditional ways for Parkinson’s patients. There is more evidence showing that suitable exercise can be a very helpful adjuvant therapy for Parkinson’s patients. Exercise can slow the disease and contribute to social well-being and mental health.

There are some creative applications in the phone or website that focus on the exercise for the Parkinson’s patients, like PD Warrior, Daily Dose PD. However, these applications may not consider the state of illness of the patients and cannot promote patients to exercise more.

In this dissertation, I present a brand-new idea about how to help Parkinson’s patients exercise effectively and do more exercise. The idea is implemented by iOS application. The core idea behind the application is the self-defined recommended algorithm. This algorithm combines the state of patients’ illness and the feature of the great exercise videos. The ultimate aim is to recommend the suitable videos to the Parkinson’s patients.

About the Speaker:

Yang Yuxiang is currently a full-time MSc(CS) student of the Department of Computer Science in the University of Hong Kong. His supervisor is Dr. Chuan Wu.