Cover Image

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

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Chen, Yiwei, Yao, Yuguang, Zhang, Yihua, Shen, Bingquan, Liu, Gaowen, Liu, Sijia
Format:	Preprint
Published:	2025
Subjects:	Artificial Intelligence Machine Learning
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency
by: Pal, Soumyadeep, et al.
Published: (2024)

Unlearners Can Lie: Evaluating and Improving Honesty in LLM Unlearning
by: Gu, Renjie, et al.
Published: (2026)

SEUF: Is Unlearning One Expert Enough for Mixture-of-Experts LLMs?
by: Zhuang, Haomin, et al.
Published: (2024)

Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning
by: Shang, Bingqi, et al.
Published: (2025)

RaVL: Discovering and Mitigating Spurious Correlations in Fine-Tuned Vision-Language Models
by: Varma, Maya, et al.
Published: (2024)

When Bias Pretends to Be Truth: How Spurious Correlations Undermine Hallucination Detection in LLMs
by: Wang, Shaowen, et al.
Published: (2025)

Superclass-Guided Representation Disentanglement for Spurious Correlation Mitigation
by: Liu, Chenruo, et al.
Published: (2025)

Model Sparsity Can Simplify Machine Unlearning
by: Jia, Jinghan, et al.
Published: (2023)

SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
by: Fan, Chongyu, et al.
Published: (2023)

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs
by: Panfilov, Alexander, et al.
Published: (2025)

The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning
by: Shah, Raj Sanjay, et al.
Published: (2026)

Can Vision Models Truly Forget? Mirage: Representation-Level Certification of Visual Unlearning
by: Yu, Zhenyu, et al.
Published: (2026)

Test-Time Training Undermines Safety Guardrails
by: Antonelli, Simone, et al.
Published: (2026)

Understanding and Preserving Safety in Fine-Tuned LLMs
by: Zhang, Jiawen, et al.
Published: (2026)

Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference
by: Ji, Jiabao, et al.
Published: (2024)

Mitigating Spurious Correlations via Disagreement Probability
by: Han, Hyeonggeun, et al.
Published: (2024)

Adversarial Watermarking for Face Recognition
by: Yao, Yuguang, et al.
Published: (2024)

Open Problems in Machine Unlearning for AI Safety
by: Barez, Fazl, et al.
Published: (2025)

The Pragmatic Frames of Spurious Correlations in Machine Learning: Interpreting How and Why They Matter
by: Bell, Samuel J., et al.
Published: (2024)

Invisible Watermarks, Visible Gains: Steering Machine Unlearning with Bi-Level Watermarking Design
by: Sun, Yuhao, et al.
Published: (2025)

The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety
by: Springer, Max, et al.
Published: (2026)

Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance
by: Zhang, Jiawen, et al.
Published: (2026)

An Adversarial Perspective on Machine Unlearning for AI Safety
by: Łucki, Jakub, et al.
Published: (2024)

UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models
by: Zhang, Yihua, et al.
Published: (2024)

Machine Mirages: Defining the Undefined
by: Tembine, Hamidou
Published: (2025)

Safety-Aware Fine-Tuning of Large Language Models
by: Choi, Hyeong Kyu, et al.
Published: (2024)

Self-Mined Hardness for Safety Fine-Tuning
by: Gupta, Prakhar, et al.
Published: (2026)

Relationship-Aware Safety Unlearning for Multimodal LLMs
by: Anilkumar, Vishnu Narayanan, et al.
Published: (2026)

The Trap of Trajectory: Towards Understanding and Mitigating Spurious Correlations in Agentic Memory
by: Tang, Luoxi, et al.
Published: (2026)

Challenging Forgets: Unveiling the Worst-Case Forget Sets in Machine Unlearning
by: Fan, Chongyu, et al.
Published: (2024)

Who is in the Spotlight: The Hidden Bias Undermining Multimodal Retrieval-Augmented Generation
by: Yao, Jiayu, et al.
Published: (2025)

Narrow Fine-Tuning Erodes Safety Alignment in Vision-Language Agents
by: Gulati, Idhant, et al.
Published: (2026)

Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment
by: Xue, Zhiyu, et al.
Published: (2026)

No, of Course I Can! Deeper Fine-Tuning Attacks That Bypass Token-Level Safety Mechanisms
by: Kazdan, Joshua, et al.
Published: (2025)

Re-Emergent Misalignment: How Narrow Fine-Tuning Erodes Safety Alignment in LLMs
by: Giordani, Jeremiah
Published: (2025)

Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models
by: Tan, Rui Yang, et al.
Published: (2026)

How Hard Can It Be? Hardness-Aware Multi-Objective Unlearning
by: Chen, Jiangwei, et al.
Published: (2026)

Reasoning Model Unlearning: Forgetting Traces, Not Just Answers, While Preserving Reasoning Skills
by: Wang, Changsheng, et al.
Published: (2025)

Does Machine Unlearning Preserve Clinical Safety? A Risk Analysis for Medical Image Classification
by: Falcao, Andreza M. C., et al.
Published: (2026)

Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning
by: Wang, Changsheng, et al.
Published: (2025)