Leyi Sheng

he/him

Guangzhou, China

I am an undergraduate student at the Hong Kong University of Science and Technology (Guangzhou), working at the intersection of AI safety and AI for safety.

My research focuses on making AI systems more robust and secure, and on using AI to advance safety in practice.

Research interests

AI for safety AI safety Image protection & steganalysis Multimodal security

Selected research

Text-to-image safety · 2026

PixJail

A self-evolving framework that turns jailbreak papers into reproducible evaluation pipelines for text-to-image models.

Image protection · 2026

IP-Bench

A systematic benchmark for evaluating image-protection methods in image-to-video generation.

Steganography security · 2026

Stego Battlefield

An evaluation of image steganography attacks and steganalysis defenses under practical security scenarios.

Offline Minecraft agent research built on pretrained world models, combining VPT skill-conditioned behavior cloning, calibrated rewards, imagination-based reinforcement learning, and world-model adaptation.

TikTok Link Scraper

A Playwright-based tool for collecting video links and metadata from Douyin, TikTok, and YouTube Shorts, with JSON output and optional video downloads via yt-dlp.

selected publications

PixJail: Self-Evolving Paper-to-Pipeline Reproduction for Text-to-Image Jailbreak Evaluation

Leyi Sheng, Han Sun, Zhen Sun, Yuntao Yue, Jinlin Wu, Xinlei He, and Jiaheng Wei

arXiv preprint, Jun 2026

Abs arXiv Bib HTML

As text-to-image jailbreak techniques evolve rapidly, existing benchmarks and reproduction workflows often struggle to keep pace. PixJail is a self-evolving paper-to-pipeline agent framework for reproducible text-to-image jailbreak evaluation. It constructs paper-specific attack modules and runnable evaluation pipelines under a unified contract while maintaining reusable, versioned reproduction knowledge.
@article{sheng2026pixjail, title = {PixJail: Self-Evolving Paper-to-Pipeline Reproduction for Text-to-Image Jailbreak Evaluation}, author = {Sheng, Leyi and Sun, Han and Sun, Zhen and Yue, Yuntao and Wu, Jinlin and He, Xinlei and Wei, Jiaheng}, journal = {arXiv preprint}, year = {2026}, month = jun, }
Stego Battlefield: Evaluating Image Steganography Attacks and Steganalysis Defenses

Zhen Sun, Zongmin Zhang, Leyi Sheng, Yule Liu, Yifan Liao, Ke Li, Xinhu Zheng, Jiaheng Wei, Wenyuan Yang, and Xinlei He

arXiv preprint, May 2026

Abs arXiv Bib HTML

SADBench systematically evaluates image steganography attacks and steganalysis defenses under practical security scenarios. It covers attack capability, defense capability, efficiency, and transferability across image-payload and text-payload steganography.
@article{sun2026stego, title = {Stego Battlefield: Evaluating Image Steganography Attacks and Steganalysis Defenses}, author = {Sun, Zhen and Zhang, Zongmin and Sheng, Leyi and Liu, Yule and Liao, Yifan and Li, Ke and Zheng, Xinhu and Wei, Jiaheng and Yang, Wenyuan and He, Xinlei}, journal = {arXiv preprint}, year = {2026}, month = may, }
IP-Bench: Benchmark for Image Protection Methods in Image-to-Video Generation Scenarios

Xiaofeng Li, Leyi Sheng, Zhen Sun, Zongmin Zhang, Jiaheng Wei, and Xinlei He

arXiv preprint, Apr 2026

Abs arXiv Bib HTML

IP-Bench is a systematic benchmark for evaluating image protection methods in image-to-video generation scenarios. It evaluates representative protection methods and state-of-the-art image-to-video models, including robustness, transferability, and practical preprocessing attacks.
@article{li2026ipbench, title = {IP-Bench: Benchmark for Image Protection Methods in Image-to-Video Generation Scenarios}, author = {Li, Xiaofeng and Sheng, Leyi and Sun, Zhen and Zhang, Zongmin and Wei, Jiaheng and He, Xinlei}, journal = {arXiv preprint}, year = {2026}, month = apr, }