Research
Research Interests
- AI Cybersecurity: AI Watermarks, LLM Security and Privacy
- Software Security: Patch Management, Vulnerability Analysis, Automated Program Repair
- Biometric Security: Automatic Speech Recognition, Computer Vision and Image Processing
Research Projects
Ph.D. Projects
Secure Voice Processing Systems against Malicious Voice Attacks
Advisors: Dr. Kun Sun
Automatic speech recognition (ASR) systems (such as Amazon Alexa and Apple Siri) are some of the widely-used human-computer interaction systems that provide convenient voice-controlled services to users. However, ASR systems are vulnerable to adversarial audio attacks that are performed by experienced attackers with modern signal processing techniques. Our research focuses on analyzing the potential vulnerabilities of automatic speech recognition systems and designing the corresponding countermeasures against adversarial audio attacks. Currently, ASR systems identify adversarial audio (e.g., replay audio) by utilizing the methods based on the frequency spectrum features. However, there still exists an arms race between attackers and defenders, with attackers developing more effective methods to evade the detection model. The frequency-based defenses may leave potential vulnerabilities if attackers are able to generate more realistic audio by frequency compensation technique. Therefore, it is crucial to understand the vulnerabilities of automatic speech recognition systems and develop more effective defenses against advanced audio attacks.
- ModReplay: modulated replay attacks with mitigation methods.
- SIEVE: secure in-vehicle ASR.
ACE: acoustic compensation emulation system.
Software Vulnerability Analysis and Security Patch Identification
Advisors: Dr. Kun Sun
Open source software (OSS) has been widely used in both free and proprietary applications. The Black Duck reports that 96% of their scanned applications contain open source components, which account for 57% of the code base on average. At the same time, vulnerabilities embedded in upstream OSS are fast propagated to the underlying applications. Also, the clone or reuse of OSS without explicit reference makes it challenging for maintainers to track and mitigate vulnerabilities. Our research develops practical techniques for detecting such vulnerabilities, which help build a more reliable and secure information system infrastructure.
- PatchDB: a large-scale security patch dataset.
- PatchRNN: a sequential-based model for security patch detection.
- GraphSPD: a graph-based model for security patch detection.
Master’s Projects
Vehicle Detection and Recognition based on Deep Neural Networks
Advisor: Dr. Feng Liu
In this project, we first built a vehicle detection system by transfer learning over the Faster-RCNN model. With the feature maps in VGG network and the weak labels in a pre-training extreme learning machine, we then built an adaptive clustering algorithm to classify the vehicle types in ever-changing scenarios. Finally, we designed a robust ELM classifier to identify the vehicle manufacturers and vehicle models.
3D Facial Image Recognition System based on Microsoft Kinect
Advisor: Dr. Feng Liu
In this project, we designed a 3D face recognition system by using Kinect as an input device. The RGB images, deep images, and facial landmarks are all collected as features with Principal Component Analysis (PCA). The final recognition system is based on SVM and written in C++ and OpenCV.
