eat, drink, and find something interesting ...

• handled computer security & privacy issues using AI; also researched the security of AI.• utilized static analysis, NLP, and graph learning to identify security patches in the OSS.• uncovered vulnerabilities in audio processing systems and mitigated the attack surfaces.• emulated user behaviors to improve honeypot fidelity & defeat sandbox-evading malware.[Software Security]• released a large-scale patch dataset PatchDB from the NVD and open-source software.• identified security patches on code revision and commit messages with PatchRNN model.• represented software patches as code property graphs and built the GraphSPD model.• applied large language models to summarize how security patches fix the vulnerabilities.[Biometric Security - Audio]• uncovered the modulated replay attack that can bypass all frequency-based detection.• designed replay detection system DualGuard with time- & frequency-domain features.• built a system prototype of secure in-vehicle ASR - SIEVE to identify command issuers.• devised an acoustic compensation emulator ACE to defeat spectrum reduction attacks.[AI Security]• penetration testing on third-party AI vendors to ensure data boundary.• secure and robust AI model watermarking with cryptographic chains.

• large language model-related security/privacy applications (data boundary assurance).• designed a system prototype; filed an utility patent; a paper accepted by ACM CCS'24.

• conducted research on cross-camera domain adaptation for face recognition task.• designed algorithm to disentangle faces to content, device, environment components.• improved cross-domain recognition accuracy via data augmentation and style transfer.• verified the feasibility of proposed algorithm by training deep learning models on AWS.• wrote code, submitted code reviews, wrote technical docs/paper, made presentations.

• researched computer vision and image processing, including image super-resolution, 3D face recognition, vehicle detection, vehicle type recognition, and license plate recognition.• implemented convolutional neural network models in GPU environments with C++/Python/Matlab language, OpenCV/VLFeat library, and Caffe framework.[Vehicle Recognition]• improved the vehicle detection rate to 98% and vehicle type recognition rate by 2.76%.• applied adaptive clustering and deep convolutional neural networks to classify vehicles.[Face Recognition]• developed a 3D face recognition system that identifies persons with 95% precision rate.• recognized faces by RGB features, deep maps, and Candide-3 landmarks with the PCA.