Sparse Autoencoders as Plug-and-Play Firewalls for Adversarial Attack Detection in VLMs
Authors: Hao Wang, Yiqun Sun, Pengfei Wei, Lawrence B. Hsieh, Daisuke Kawahara
Organizations: 2Waseda University · 1Magellan Technology Research Institute (MTRI)
Abstract
Vision-language models (VLMs) have advanced rapidly and are increasingly deployed in real-world applications, especially with the rise of agent-based systems. However, their safety has received relatively limited attention. Even the latest proprietary and open-weight VLMs remain highly vulnerable to adversarial attacks, leaving downstream applications exposed to significant risks. In this work, we propose a novel and lightweight adversarial attack detection framework based on sparse autoencoders (SAEs), termed SAEgis. By inserting an SAE module into a pretrained VLM and training it with standard reconstruction objectives, we find that the learned sparse latent features naturally capture attack-relevant signals. These features enable reliable classification of whether an input image has been adversarially perturbed, even for previously unseen samples. Extensive experiments show that SAEgis achieves strong performance across in-domain, cross-domain, and cross-attack settings, with particularly large improvements in cross-domain generalization compared to existing baselines. In addition, combining signals from multiple layers further improves robustness and stability. To the best of our knowledge, this is the first work to explore SAE as a plug-and-play mechanism for adversarial attack detection in VLMs. Our method requires no additional adversarial training, introduces minimal overhead, and provides a practical approach for improving the safety of real-world VLM systems.
Targeted adversarial attacks on Large Vision-Language Models (LVLMs) test whether small image perturbations can steer model responses toward attacker-specified content. Under the standard L-infinity constraint, targeted attacks become a regional perturbation budget allocation problem: attack success depends not only on the perturbation objective, but also on which regions receive updates and in what order. Existing localized attacks improve over global perturbations but rely on stochastic spatial sampling, often updating weakly influential regions. We address this limitation through an attention-based analysis showing that cross-modal attention identifies adversarially sensitive regions and that perturbing high-attention hotspots induces predictable redistribution toward subsequent salient regions. These findings motivate attention-guided region sequencing, which begins from dominant hotspots and progressively moves the update support toward next-salient regions. Based on these principles, we propose Stage-wise Attention-Guided Attack (SAGA), a black-box region-sequencing framework that uses a fixed attention map from an open-source LVLM to guide perturbation updates without accessing target-model parameters, gradients, or attention maps. Across ten closed-source and open-source LVLMs, SAGA achieves state-of-the-art attack success rates and the best overall imperceptibility. The source code is available at https://github.com/jaehyun-kwak/SAGA.
Vision-language pre-trained models (VLPs) are widely used in real-world applications. However, they remain vulnerable to adversarial attacks. Although adversarial detection methods have demonstrated success in single-modality settings (either vision or language), their effectiveness and reliability in multimodal models such as VLPs remain largely unexplored. In this work, we study the geometry of VLP embedding spaces and observe structured anisotropy that differs from unimodal vision models. Our theoretical analysis shows that under this anisotropic structure, adversarial attacks increase the expected geometric separation between clean and adversarial examples (AEs). Specifically, we demonstrate that AEs consistently exhibit greater expected distances to randomly sampled points than their clean counterparts, indicating that AEs tend to push representations out of manifold regions. Building on these insights, we propose GeoDetect, which leverages these off-manifold deviations via geometric scores to identify AEs. Through comprehensive evaluations, we show that our approach reliably detects AEs across diverse VLP architectures and threat settings, covering unimodal and multimodal attacks as well as adaptive attacks, thereby providing a robust and practical approach to improving the safety and reliability of these models.
Afsaneh Hasanebrahimi, Hanxun Huang, Christopher Leckie +2
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.