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

arXiv: https://arxiv.org/abs/2503.11832 · HTML: https://mlsanity.com/paper/2503.11832/

## Key facts

- Cited by 16 papers in the map (data through 14 Oct 2026).
- Placed in MLLM jailbreaks and adversarial robustness, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 81 of its 92 references that are in the map.
- First posted on arXiv 14 Mar 2025.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM jailbreaks and adversarial robustness](https://mlsanity.com/mllm/mllm-jailbreaks-and-adversarial-robustness/)
- LLM safety and security / [Vision-language and text-to-image safety](https://mlsanity.com/safety/vision-language-and-text-to-image-safety/)

## Builds on (most cited first, 30 of 81)

- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://mlsanity.com/paper/2301.12597/) (2023)
- [GPT-4o System Card](https://mlsanity.com/paper/2410.21276/) (2024)
- [Improved Baselines with Visual Instruction Tuning](https://mlsanity.com/paper/2310.03744/) (2023)
- [Flamingo: a Visual Language Model for Few-Shot Learning](https://mlsanity.com/paper/2204.14198/) (2022)
- [Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution](https://mlsanity.com/paper/2409.12191/) (2024)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models](https://mlsanity.com/paper/2304.10592/) (2023)
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://mlsanity.com/paper/1612.00837/) (2016)
- [Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering](https://mlsanity.com/paper/2209.09513/) (2022)
- [Towards VQA Models That Can Read](https://mlsanity.com/paper/1904.08920/) (2019)
- [Jailbroken: How Does LLM Safety Training Fail?](https://mlsanity.com/paper/2307.02483/) (2023)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [Machine Unlearning](https://mlsanity.com/paper/1912.03817/) (2019)
- [mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality](https://mlsanity.com/paper/2304.14178/) (2023)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [VizWiz Grand Challenge: Answering Visual Questions from Blind People](https://mlsanity.com/paper/1802.08218/) (2018)
- [MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning](https://mlsanity.com/paper/2310.09478/) (2023)
- [LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model](https://mlsanity.com/paper/2304.15010/) (2023)
- [OpenFlamingo: An Open-Source Framework for Training Large Autoregressive Vision-Language Models](https://mlsanity.com/paper/2308.01390/) (2023)
- [The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning](https://mlsanity.com/paper/2403.03218/) (2024)
- [XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models](https://mlsanity.com/paper/2308.01263/) (2023)
- [Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning](https://mlsanity.com/paper/2404.05868/) (2024)
- [VisionLLM: Large Language Model is also an Open-Ended Decoder for Vision-Centric Tasks](https://mlsanity.com/paper/2305.11175/) (2023)
- [Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation](https://mlsanity.com/paper/2310.06987/) (2023)
- [Jailbreak and Guard Aligned Language Models with Only Few In-Context Demonstrations](https://mlsanity.com/paper/2310.06387/) (2023)
- [FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts](https://mlsanity.com/paper/2311.05608/) (2023)
- [Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions](https://mlsanity.com/paper/2309.07875/) (2023)

## Built on by (most cited first, 16 of 16)

- [LLM Unlearning Reveals a Stronger-Than-Expected Coreset Effect in Current Benchmarks](https://mlsanity.com/paper/2504.10185/) (2025)
- [Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs](https://mlsanity.com/paper/2506.14003/) (2025)
- [Unleashing the Power of Chain-of-Prediction for Monocular 3D Object Detection](https://mlsanity.com/paper/2505.04594/) (2025)
- [Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks](https://mlsanity.com/paper/2607.07907/) (2026)
- [Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning](https://mlsanity.com/paper/2510.17021/) (2025)
- [Towards Intrinsic-Aware Monocular 3D Object Detection](https://mlsanity.com/paper/2603.27059/) (2026)
- [Concept Regions Matter: Benchmarking CLIP with a New Cluster-Importance Approach](https://mlsanity.com/paper/2511.12978/) (2025)
- [Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images](https://mlsanity.com/paper/2603.08486/) (2026)
- [LLM Unlearning on Noisy Forget Sets: A Study of Incomplete, Rewritten, and Watermarked Data](https://mlsanity.com/paper/2510.09007/) (2025)
- [Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck](https://mlsanity.com/paper/2603.10351/) (2026)
- [Rotation Control Unlearning: Quantifying and Controlling Continuous Unlearning for LLM with The Cognitive Rotation Space](https://mlsanity.com/paper/2509.25743/) (2025)
- [Supervised Fine-Tuning Needs to Unlock the Potential of Token Priority](https://mlsanity.com/paper/2602.01227/) (2026)
- [Shortcuts in the Tail: Debiasing via Post-Hoc Spectral Compression of Fine-Tuning Updates](https://mlsanity.com/paper/2606.07596/) (2026)
- [Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space](https://mlsanity.com/paper/2608.07786/) (2026)
- [Partition-Aware Unlearning for Removing Spurious Correlations in Large Vision-Language Models](https://mlsanity.com/paper/2608.29996/) (2026)
- [A2Safe: Counterfactual Evidence-Aligned Adaptive Agent Collaboration for Safe and Effective Visual Question Answering](https://mlsanity.com/paper/2609.24098/) (2026)

## Abstract (arXiv)

Recent vision language models (VLMs) have made remarkable strides in generative modeling with multimodal inputs, particularly text and images. However, their susceptibility to generating harmful content when exposed to unsafe queries raises critical safety concerns. While current alignment strategies primarily rely on supervised safety fine-tuning with curated datasets, we identify a fundamental limitation we call the ''safety mirage'', where supervised fine-tuning inadvertently reinforces spurious correlations between superficial textual patterns and safety responses, rather than fostering deep, intrinsic mitigation of harm. We show that these spurious correlations leave fine-tuned VLMs vulnerable even to a simple one-word modification-based attack, where substituting a single word in text queries with a spurious correlation-inducing alternative can effectively bypass safeguards. Additionally, these correlations contribute to the over-prudence, causing fine-tuned VLMs to refuse benign queries unnecessarily. To address these issues, we show machine unlearning (MU) as a powerful alternative to supervised safety fine-tuning, as it avoids biased feature-label mappings and directly removes harmful knowledge from VLMs while preserving their general capabilities. Extensive evaluations across safety benchmarks show that under MU-based alignment reduces the attack success rate by up to 60.27% and cuts unnecessary rejections by over 84.20%. WARNING: There exist AI generations that may be offensive in nature.