Membership Inference Attacks Against Video Large Language Models
SiMLA 2026
Wei Song
Postdoctoral Research Fellow, UNSW Sydney
I am a postdoctoral researcher at UNSW Sydney, supervised by Prof. Jingling Xue, working on the security, reliability, and real-world deployment of AI-enabled systems. My research focuses on AI systems security, trustworthy media, adversarial robustness, and multimodal model safety, especially for video intelligence and large language model ecosystems.
I collaborate with academic and industry partners to build practical technologies for content authenticity and model trustworthiness. Current research focuses on security and safety challenges in AI-generated media, robust multimodal reasoning, and secure AI-powered infrastructure for critical domains such as energy systems.
I have taught across 22 terms at UNSW with substantial responsibilities in tutorials, laboratories, assessment design, and course administration, and have consistently received positive feedback from both students and course convenors.
Research Interests
- AI Systems Security
- Trustworthy Media
- Multimodal AI Safety
- Adversarial Video Attacks and Defenses
- Watermarking for AI-Generated Content
- Secure AI-Powered Energy Systems
Highlights
- Apr 2026: Two papers accepted at ACL 2026.
- Mar 2026: Paper accepted at ICME 2026.
- Jan 2026: NgCaptcha accepted at WWW 2026 (Demo).
- Nov 2025: VideoLLM poisoning paper accepted at AAAI 2026.
- Jul 2025: ViDToken accepted at RAID 2025.
- Jun 2025: HandID accepted at IMWUT/UbiComp 2025.
- May 2025: SynerGuard accepted at IEEE ICRA 2025.
- Jun 2024: First AI adversarial attack/defense work accepted at USENIX Security 2024.
- Mar 2024: VibMilk accepted at IEEE Internet of Things Journal 2024.
- Jan 2024: CAvatar accepted at IMWUT/UbiComp 2024.
Publications
Train in Vain: Functionality-Preserving Poisoning to Prevent Unauthorized Use of Code Datasets
ACL Findings 2026
Debiasing LLMs by Masking Unfairness-Driving Attention Heads
ACL Findings 2026 · Paper
DUAP: Dual-task Universal Adversarial Perturbations Against Voice Control Systems
ICME 2026 · Paper
NgCaptcha: A CAPTCHA Bridging the Past and the Future
WWW 2026 (Demo) · Paper
DeMark: A Query-Free Black-Box Attack on Deepfake Watermarking Defenses
arXiv 2026 · Paper
VideoSTF: Stress-Testing Output Repetition in Video Large Language Models
arXiv 2026 · Paper
ViDToken: A Video-Transformer-Based Latent Token Defense for Adversarial Video Detection
RAID 2025 · Paper
Failures to Surface Harmful Contents in Video Large Language Models
AAAI 2026 · Paper
Poisoning Prompt-Guided Sampling in Video Large Language Models
arXiv · Paper
Help or Hurdle? Rethinking Model Context Protocol-Augmented Large Language Models
arXiv · Paper
SynerGuard: A Robust Framework for Point Cloud Classification via Local Geometry and Spatial Topology
IEEE ICRA 2025 · Paper
Correction-based Defense Against Adversarial Video Attacks via Discretization-Enhanced Video Compressive Sensing
USENIX Security 2024 · Paper
HandID: Unobtrusive Gesture-independent User Authentication on Smartphones Using Vibration-based Hand Biometrics
ACM IMWUT/UbiComp 2025 · Paper
VibMilk: Non-intrusive Milk Spoilage Detection via Smartphone Vibration
IEEE Internet of Things Journal 2024 · Paper
CAvatar: Real-time Human Activity Mesh Reconstruction via Tactile Carpets
ACM IMWUT/UbiComp 2024 · Paper
Pistis: Replay Attack and Liveness Detection for Gait-Based User Authentication System on Wearable Devices Using Vibration
IEEE Internet of Things Journal 2022 · Paper
Poster: Towards Behavior-independent In-hand User Authentication on Smartphone Using Vibration
ACM MobiCom 2022 Poster · Paper
Experience
- Postdoctoral Research Fellow, UNSW Sydney (Nov 2024 - Present). AI-enabled systems security, trustworthy media, multimodal safety, and collaboration with Data61.
- Tutor / Course Admin, UNSW Sydney (2019 - 2024). 22 terms across computer networks, IoT, mobile networking, wireless security, and deep learning.
- Software Engineer, Parking Spotz (Jan 2022 - May 2022). Mobile application design and implementation for intelligent parking systems.
- Research Assistant, UNSW/CSCRC (Dec 2020 - May 2021). Developed secure and privacy-preserving mobile prototype systems.
Professional Service
- Reviewer for ICLR, ICML, CVPR, AAAI, FSE, ASE, SIGCOMM, MobiSys, IoTDI, and IPSN.
- Artifact Evaluation Committee member for USENIX Security.