cs.CVNov 16, 2026

SmoothShift: Disrupting Visual Adversarial Attacks on LVLMs at Test Time

Authors: Xiaoqing Chen, Jiayi Ouyang, Dongrui Wu

Organizations: Huazhong University of Science and Technology · Zhongguancun Academy

Abstract

Visual adversarial perturbations can substantially impair large vision-language models (LVLMs) despite introducing only subtle input changes. To understand how these small changes cause substantial errors, we investigate their optimization and propagation through the model, revealing two key findings. First, stronger attacks rely on finer iterative optimization, while subtle input perturbations induce representation discrepancies that are amplified layer by layer through the vision encoder and persist in the language model. Second, disrupting the alignment between within-patch input patterns and the model's computation reduces clean-adversarial representation discrepancies and substantially weakens attacks. Together, these findings suggest that adversarial effectiveness depends on both fine-grained perturbation patterns and their coupling with the model's computation. Motivated by these findings, we propose SmoothShift, a lightweight, training-free defense that combines local smoothing and spatial transformations. Smoothing weakens fine-grained perturbation patterns, while spatial transformations disrupt their alignment with the encoder's patch grid. SmoothShift requires no model updates, auxiliary networks, or additional model inference. Experiments cover three LVLM families on four datasets and robustly fine-tuned Contrastive Language-Image Pre-training (CLIP) models on eight classification datasets. Evaluations across eight attack methods and multiple perturbation budgets demonstrate improved robustness while largely preserving clean performance. SmoothShift also provides robustness gains under defense-aware attacks using mixed-gradient BPDA with EOT.

Figures & tables

Explore similar work

Apr 1, 2026

PDA: Text-Augmented Defense Framework for Robust Vision-Language Models against Adversarial Image Attacks

Vision-language models (VLMs) are vulnerable to adversarial image perturbations. Existing works based on adversarial training against task-specific adversarial examples are computationally expensive and often fail to generalize to unseen attack types. To address these limitations, we introduce Paraphrase-Decomposition-Aggregation (PDA), a training-free defense framework that leverages text augmentation to enhance VLM robustness under diverse adversarial image attacks. PDA performs prompt paraphrasing, question decomposition, and consistency aggregation entirely at test time, thus requiring no modification on the underlying models. To balance robustness and efficiency, we instantiate PDA as invariants that reduce the inference cost while retaining most of its robustness gains. Experiments on multiple VLM architectures and benchmarks for visual question answering, classification, and captioning show that PDA achieves consistent robustness gains against various adversarial perturbations while maintaining competitive clean accuracy, establishing a generic, strong and practical defense framework for VLMs during inference.
Dec 11, 2024

Exploiting Vision Encoder Vulnerabilities for Universal Adversarial Perturbations on Large Vision-Language Models

Large Vision-Language Models (LVLMs) have achieved remarkable performance on multimodal tasks but remain highly vulnerable to small adversarial perturbations in input images. Existing attacks typically target the vision encoder's final output embeddings, implicitly treating the encoder as a uniform attack surface, while a systematic analysis of which internal components are most vulnerable has remained largely unexplored. We show such analysis is essential, as adversarial vulnerability in LVLM vision encoders is structurally concentrated rather than uniformly distributed. Building on this, we propose Vision Encoder Vulnerable-Component-Targeted Universal Adversarial Perturbation (VEV-UAP), a task-agnostic and cost-efficient attack framework. Through a component- and layer-wise analysis of attention mechanisms, we identify the value components in middle layers as critical vulnerabilities that strongly influence downstream language model behavior. VEV-UAP selectively targets these components to generate a single universal perturbation shared across images, without involving textual inputs or the language model during optimization. Experiments across multiple LVLMs and tasks show VEV-UAP achieves state-of-the-art attack success rates with reduced computational overhead. Moreover, a single VEV-UAP transfers across LVLMs sharing the same vision encoder, even when paired with different language models, making it a practical framework for scalable robustness evaluation.
Feb 23, 2026

VALD: Multi-Stage Vision Attack Detection for Efficient LVLM Defense

Large Vision-Language Models (LVLMs) can be vulnerable to adversarial images that subtly bias their outputs toward plausible yet incorrect responses. We introduce a general, efficient, and training-free defense that combines image transformations with agentic data consolidation to recover correct model behavior. A key component of our approach is a two-stage detection mechanism that quickly filters out the majority of clean inputs. We first assess image consistency under content-preserving transformations at negligible computational cost. For more challenging cases, we examine discrepancies in a text-embedding space. Only when necessary do we invoke a powerful LLM to resolve attack-induced divergences. A key idea is to consolidate multiple responses, leveraging both their similarities and their differences. We show that our method achieves state-of-the-art accuracy while maintaining notable efficiency: most clean images skip costly processing, and even in the presence of numerous adversarial examples, the overhead remains minimal.