Auditing Privacy Leakage in Tabular Foundation Model Embeddings
Cascade Probing audits sensitive-attribute leakage from tabular foundation-model embeddings and reveals a privacy cliff when additional attributes are known.
Publications - Xun Wang
You can also find my articles on my Google Scholar profile.
AgentProv: Auditing Agentic LLM API Providers via Tool-use Policy Probes
AgentProv audits the identity of agentic LLM APIs through their tool-selection behavior, using action-based fingerprints and an MMD permutation test.
Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models
A systematic benchmark investigating privacy risks under differentially private adaptations of LLMs using state-of-the-art attacks such as robust membership inference and canary data extraction.
Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs
A novel framework for transferring soft prompts between LLMs while preserving privacy through differential privacy and knowledge distillation techniques.
Provenance Tracing for World Models: Was this Agent Trained in my World Model?
Provenance tracing for agents trained in world models.
Frequency-Domain Model Fingerprinting for Image Autoregressive Models
FreqIAR: A frequency-domain approach for fingerprinting image autoregressive models.
Structurally Prune Anything: Any Architecture, Any Framework, Any Time
A versatile structured pruning framework that works across any neural network architecture, framework, and training stage, introducing OBSPA for post-training pruning without fine-tuning.