Sijia Liu

Associate Professor@Michigan State University·美国·OpenReview

研究方向

Machine Unlearning · Model pruning · Adversarial Machine Learning · Trustworthy AI · Zeroth-order optimization · Optimization for machine learning · Sparse learning

Sijia Liu

Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-Tuning and Can Be Mitigated by Machine Unlearning

Attention Smoothing Is All You Need For Unlearning

Old Memories Die Hard: Understanding Challenges of Privacy Unlearning in Large Language Models

CyclicReflex: Improving Reasoning Models via Cyclical Reflection Token Scheduling

Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs

Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning

Forget Vectors at Play: Universal Input Perturbations Driving Machine Unlearning in Image Classification

Label Smoothing Improves Machine Unlearning

Tackling Fake Forgetting through Uncertainty Quantification

The Dual Power of Interpretable Token Embeddings: Jailbreaking Attacks and Defenses for Diffusion Model Unlearning

Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding

(Be Cautious!) Bio-Foundation Models Are Not Yet Robust to Biological Plausible Perturbations and ML Transformations

LLM Unlearning Under the Microscope: A Full-Stack View on Methods and Metrics

EPiC: Towards Lossless Speedup for Reasoning Training through Edge-Preserving CoT Condensation

FRELA: Frequency-Layered Audio Watermarking for Robust Content Authentication

When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers

The Fragile Truth of Saliency: Improving LLM Input Attribution via Attention Bias Optimization

Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning

Forget Vectors at Play: Universal Input Perturbations Driving Machine Unlearning in Image Classification

Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond

One Token Embedding Is Enough to Deadlock Your Large Reasoning Model

Scalable Back-Propagation-Free Training of Optical Physics-Informed Neural Networks

Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning

LLM Unlearning Reveals a Stronger-Than-Expected Coreset Effect in Current Benchmarks

UOE: Unlearning One Expert is Enough for Mixture-of-Experts LLMs

Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning

Forget Vectors at Play: Universal Input Perturbations Driving Machine Unlearning in Image Classification

Visual Prompting Reimagined: The Power of Activation Prompts

FairSkin: Fair Diffusion for Skin Disease Image Geneartion