Android Infinite Loop Detection Based on Static Analysis (collaboration with Huawei)

Supervisor: Yepang Liu
  • Modeled Android programs into graphical representations such as control flow graph, module call graph, method call graph.
  • Identified natural loops from graphical code representations.
  • Conducted static analysis on natural loop snippets to identify loop termination conditions based on Loopster algorithm.

Multi-Criteria Test-Suite Minimization with Reinforcement Learning (RL)

Supervisor: Joshua Garcia
  • Efficiently and effectively solved MCTSM problem through mapping it into a RL problem.
  • Implemented policy gradient based RL algorithms like REINFORCE and Actor-Critic.
  • Greatly exceeded the scalability and efficiency of traditional methods based on integer linear programming under a large test-suite size and a condition of multiple criteria condition.

How Do Python Framework APIs Evolves? An Exploratory Study[SANER’20]

Supervisor: Yepang Liu
  • Performed the first systematic study to characterize the evolution of Python framework APIs and discussed the similarities and differences between the API evolution in Python and Java frameworks.
  • Quantitatively and qualitatively analyzed the types of compatibility issues caused by misusing evolved APIs in Python applications and the commonly adopted fixing strategies.
  • Designed and implemented a tool, PYCOMPAT, to detect compatibility issues in Python applications.

Android Malware Evading using Reinforcement Learning

Supervisor: Joshua Garcia
  • Developed an automatic malware evading framework that leveraged reinforcement learning techniques.
  • Leveraged code obfuscation tools like Allatori, DroidChameleon, and Adam to modify Android malware to evade the detection of the state-of-the-art machine-learning based anti-malware RevealDroid.
  • Implemented Q-learning based RL algorithm (Deep Q-learning algorithm) as the agent.
  • Successfully automated the generation of malware samples that can evade detection and are on par with the performance of the state-of-art approach.