Poison as Cure: Visual Noise for Mitigating Object Hallucinations in LVMs

TL;DR

This work tackles object hallucination in large vision-language models by introducing Visual Adversarial Perturbation (VAP), a data-centric method that injects beneficial visual noise into inputs without modifying the base models. VAP optimizes three strategy losses via a zero-gradient, black-box approach to align model outputs with actual visual content while reducing dependence on biased parametric knowledge. Across eight state-of-the-art LVMs and three benchmark suites (POPE, BEAF, CHAIR), VAP consistently reduces hallucinations and improves visual grounding, with analyses detailing false drops, perturbation strength, and TU gains. The approach is architecture-agnostic and can be scaled using proxy models to mitigate computational costs, offering a practical path to more reliable multimodal systems in real-world applications.

Abstract

Large vision-language models (LVMs) extend large language models (LLMs) with visual perception capabilities, enabling them to process and interpret visual information. A major challenge compromising their reliability is object hallucination that LVMs may generate plausible but factually inaccurate information. We propose a novel visual adversarial perturbation (VAP) method to mitigate this hallucination issue. VAP alleviates LVM hallucination by applying strategically optimized visual noise without altering the base model. Our approach formulates hallucination suppression as an optimization problem, leveraging adversarial strategies to generate beneficial visual perturbations that enhance the model's factual grounding and reduce parametric knowledge bias. Extensive experimental results demonstrate that our method consistently reduces object hallucinations across 8 state-of-the-art LVMs, validating its efficacy across diverse evaluations.

Figures (8)

• Figure 1: We introduce VAP (visual adversarial perturbation), a novel approach that strategically injects beneficial visual noise to mitigate object hallucination in LVMs without altering the complex base model. Our method consistently improves performance across 8 state-of-the-art LVMs under the POPE hallucination evaluation setting pope.
• Figure 2: Detailed Overview of our proposed method. The VAP method generates beneficial visual noise by leveraging adversarial knowledge through the optimization of three strategies: (1) maximizing the semantic alignment between the LVM's response and the visual content to preserve the semantic consistency of the image, (2) minimizing the response similarity between the original and distorted visual content through noise-induced uncertainty, and (3) mitigating parametric knowledge bias by minimizing the similarity of representations between original and distorted inputs. Strategies (2) and (3) jointly mitigate parametric knowledge bias. The optimized visual noise effectively mitigates object hallucinations.
• Figure 3: Comparison of the original images with our proposed VAP and Gaussian noise of equal strength ($\epsilon=2$). We highlight the performance degradation when adding Gaussian noise compared to VAP. The experiments were conducted using eight SOTA LVMs under the POPE popular evaluation setting, with evaluations on F1 Score.
• Figure 4: Examples of the vision-question-answer (VQA) tasks before and after applying our proposed method to the original images. (a) and (b) demonstrates the suppression of hallucinations in vision-/text-axis evaluations. (c) and (d) shows the reduction of hallucinations in open-ended tasks. Specifically, we use the LLaVA-v1.5 llava as an example.
• Figure 5: Performance of the Intern-VL2 model InternVL under varying levels of perturbation strength in the POPE adversarial setting. We test the model's performance with varying perturbations applied to the original images.