Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-tuning

• URL: http://arxiv.org/abs/2503.11832v1
• Date: Fri, 14 Mar 2025 19:52:08 GMT
• Title: Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-tuning
• Authors: Yiwei Chen, Yuguang Yao, Yihua Zhang, Bingquan Shen, Gaowen Liu, Sijia Liu,
• Abstract summary: Vision-language models (VLMs) have made remarkable strides in generative modeling with multimodal inputs.<n>Current alignment strategies rely on supervised safety fine-tuning with curated datasets.<n>We show that supervised fine-tuning inadvertently reinforces spurious correlations between superficial textual patterns and safety responses.
• Score: 23.71517734919702
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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 this issue, 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 one-word attacks, MU-based alignment reduces the attack success rate by up to 60.17% and cuts unnecessary rejections by over 84.20%. Codes are available at https://github.com/OPTML-Group/VLM-Safety-MU. WARNING: There exist AI generations that may be offensive in nature.

Related papers

• Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? [83.53005932513155]
  Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. We propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences.
  arXiv  Detail & Related papers  (2025-04-14T09:03:51Z)
• Understanding and Rectifying Safety Perception Distortion in VLMs [19.239094089025095]
  Vision-language models (VLMs) become more susceptible to harmful requests and jailbreak attacks after integrating the vision modality.<n> multimodal inputs introduce an modality-induced activation shift toward a "safer" direction compared to their text-only counterparts.<n>We propose ShiftDC, a training-free method that decomposes and calibrates the modality-induced activation shift to reduce the impact of modality on safety.
  arXiv  Detail & Related papers  (2025-02-18T18:06:48Z)
• MLLM-as-a-Judge for Image Safety without Human Labeling [81.24707039432292]
  In the age of AI-generated content (AIGC), many image generation models are capable of producing harmful content. It is crucial to identify such unsafe images based on established safety rules. Existing approaches typically fine-tune MLLMs with human-labeled datasets.
  arXiv  Detail & Related papers  (2024-12-31T00:06:04Z)
• Enhancing AI Safety Through the Fusion of Low Rank Adapters [7.384556630042846]
  Low-Rank Adapter Fusion mitigates harmful responses when faced with malicious prompts.<n>We show a 42% reduction in the harmfulness rate by leveraging LoRA fusion between a task adapter and a safety adapter.<n>We also observe exaggerated safety behaviour, where the model rejects safe prompts that closely resemble unsafe ones.
  arXiv  Detail & Related papers  (2024-12-30T13:12:27Z)
• ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time [12.160713548659457]
  adversarial visual inputs can easily bypass VLM defense mechanisms.<n>We propose a novel two-phase inference-time alignment framework, evaluating input visual contents and output responses.<n>Experiments show that ETA outperforms baseline methods in terms of harmlessness, helpfulness, and efficiency.
  arXiv  Detail & Related papers  (2024-10-09T07:21:43Z)
• HiddenGuard: Fine-Grained Safe Generation with Specialized Representation Router [42.222681564769076]
  We introduce HiddenGuard, a novel framework for fine-grained, safe generation in Large Language Models. HiddenGuard incorporates Prism, which operates alongside the LLM to enable real-time, token-level detection and redaction of harmful content. Our experiments demonstrate that HiddenGuard achieves over 90% in F1 score for detecting and redacting harmful content.
  arXiv  Detail & Related papers  (2024-10-03T17:10:41Z)
• CoCA: Regaining Safety-awareness of Multimodal Large Language Models with Constitutional Calibration [90.36429361299807]
  multimodal large language models (MLLMs) have demonstrated remarkable success in engaging in conversations involving visual inputs. The integration of visual modality has introduced a unique vulnerability: the MLLM becomes susceptible to malicious visual inputs. We introduce a technique termed CoCA, which amplifies the safety-awareness of the MLLM by calibrating its output distribution.
  arXiv  Detail & Related papers  (2024-09-17T17:14:41Z)
• Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training [67.30423823744506]
  This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful
  arXiv  Detail & Related papers  (2024-07-12T09:36:33Z)
• Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patching [74.62818936088065]
  textscSafePatching is a novel framework for comprehensive PSA.<n>textscSafePatching achieves a more comprehensive PSA than baseline methods.<n>textscSafePatching demonstrates its superiority in continual PSA scenarios.
  arXiv  Detail & Related papers  (2024-05-22T16:51:07Z)
• RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content [62.685566387625975]
  Current mitigation strategies, while effective, are not resilient under adversarial attacks. This paper introduces Resilient Guardrails for Large Language Models (RigorLLM), a novel framework designed to efficiently moderate harmful and unsafe inputs.
  arXiv  Detail & Related papers  (2024-03-19T07:25:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.