Understanding and Rectifying Safety Perception Distortion in VLMs

• URL: http://arxiv.org/abs/2502.13095v1
• Date: Tue, 18 Feb 2025 18:06:48 GMT
• Title: Understanding and Rectifying Safety Perception Distortion in VLMs
• Authors: Xiaohan Zou, Jian Kang, George Kesidis, Lu Lin,
• Abstract summary: 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.
• Score: 19.239094089025095
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent studies reveal that vision-language models (VLMs) become more susceptible to harmful requests and jailbreak attacks after integrating the vision modality, exhibiting greater vulnerability than their text-only LLM backbones. To uncover the root cause of this phenomenon, we conduct an in-depth analysis and identify a key issue: multimodal inputs introduce an modality-induced activation shift toward a "safer" direction compared to their text-only counterparts, leading VLMs to systematically overestimate the safety of harmful inputs. We refer to this issue as safety perception distortion. To mitigate such distortion, we propose Activation Shift Disentanglement and Calibration (ShiftDC), a training-free method that decomposes and calibrates the modality-induced activation shift to reduce the impact of modality on safety. By isolating and removing the safety-relevant component, ShiftDC restores the inherent safety alignment of the LLM backbone while preserving the vision-language capabilities of VLMs. Empirical results demonstrate that ShiftDC significantly enhances alignment performance on safety benchmarks without impairing model utility.

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)
• Fundamental Safety-Capability Trade-offs in Fine-tuning Large Language Models [92.38300626647342]
  Fine-tuning Large Language Models (LLMs) on some task-specific datasets has been a primary use of LLMs. This paper presents a theoretical framework for understanding the interplay between safety and capability in two primary safety-aware LLM fine-tuning strategies.
  arXiv  Detail & Related papers  (2025-03-24T20:41:57Z)
• Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-tuning [23.71517734919702]
  Vision-language models (VLMs) have made remarkable strides in generative modeling with multimodal inputs. Current alignment strategies rely on supervised safety fine-tuning with curated datasets. We show that supervised fine-tuning inadvertently reinforces spurious correlations between superficial textual patterns and safety responses.
  arXiv  Detail & Related papers  (2025-03-14T19:52:08Z)
• Improving LLM Safety Alignment with Dual-Objective Optimization [65.41451412400609]
  Existing training-time safety alignment techniques for large language models (LLMs) remain vulnerable to jailbreak attacks. We propose an improved safety alignment that disentangles DPO objectives into two components: (1) robust refusal training, which encourages refusal even when partial unsafe generations are produced, and (2) targeted unlearning of harmful knowledge.
  arXiv  Detail & Related papers  (2025-03-05T18:01:05Z)
• HiddenDetect: Detecting Jailbreak Attacks against Large Vision-Language Models via Monitoring Hidden States [17.601328965546617]
  We investigate whether LVLMs inherently encode safety-relevant signals within their internal activations during inference. Our findings reveal that LVLMs exhibit distinct activation patterns when processing unsafe prompts. We introduce HiddenDetect, a novel tuning-free framework that harnesses internal model activations to enhance safety.
  arXiv  Detail & Related papers  (2025-02-20T17:14:34Z)
• VLM-Guard: Safeguarding Vision-Language Models via Fulfilling Safety Alignment Gap [51.287157951953226]
  Vision language models (VLMs) come with increased safety concerns.<n>VLMs can be built upon LLMs that have textual safety alignment, but it is easily undermined when the vision modality is integrated.<n>We propose VLM-Guard, an inference-time intervention strategy that leverages the LLM component of a VLM as supervision for the safety alignment of the VLM.
  arXiv  Detail & Related papers  (2025-02-14T08:44:43Z)
• Internal Activation as the Polar Star for Steering Unsafe LLM Behavior [50.463399903987245]
  We introduce SafeSwitch, a framework that dynamically regulates unsafe outputs by monitoring and utilizing the model's internal states.<n>Our empirical results show that SafeSwitch reduces harmful outputs by over 80% on safety benchmarks while maintaining strong utility.
  arXiv  Detail & Related papers  (2025-02-03T04:23:33Z)
• Internal Activation Revision: Safeguarding Vision Language Models Without Parameter Update [8.739132798784777]
  Vision-language models (VLMs) demonstrate strong multimodal capabilities but have been found to be more susceptible to generating harmful content.<n>We propose an textbfinternal activation revision approach that efficiently revises activations during generation.<n>Our framework incorporates revisions at both the layer and head levels, offering control over the model's generation at varying levels of granularity.
  arXiv  Detail & Related papers  (2025-01-24T06:17:22Z)
• Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models [26.83278034227966]
  The safety alignment ability of Vision-Language Models (VLMs) is prone to be degraded by the integration of the vision module. We show that the challenge arises from the representation gap that emerges when introducing vision modality to VLMs. To reduce safety alignment degradation, we introduce Cross-Modality Representation Manipulation (CMRM)
  arXiv  Detail & Related papers  (2024-10-11T17:59:31Z)
• How Does Vision-Language Adaptation Impact the Safety of Vision Language Models? [27.46416187893547]
  Vision-Language adaptation (VL adaptation) transforms Large Language Models (LLMs) into Large Vision-Language Models (LVLMs) Despite potential harmfulness due to weakened safety measures, in-depth analysis on the effects of VL adaptation on safety remains under-explored.
  arXiv  Detail & Related papers  (2024-10-10T03:12:03Z)
• 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)
• SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering [56.92068213969036]
  Safety alignment is indispensable for Large Language Models (LLMs) to defend threats from malicious instructions.<n>Recent researches reveal safety-aligned LLMs prone to reject benign queries due to the exaggerated safety issue.<n>We propose a Safety-Conscious Activation Steering (SCANS) method to mitigate the exaggerated safety concerns.
  arXiv  Detail & Related papers  (2024-08-21T10:01:34Z)

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.