QuISE: Defense against Typographic Attacks on VLMs via Query-Irrelevant Semantic Editing

Typographic attacks pose a critical threat to vision-language models (VLMs) by injecting misleading text into images and causing models to rely on adversarial textual cues rather than visual evidence. Existing defenses often require model-specific modifications, additional training, or access to internal model components, limiting their applicability to modern closed-source VLMs. In this paper, we propose QuISE, a model-agnostic, training-free black-box defense based on query-irrelevant semantic editing. QuISE first identifies text regions likely to affect the current query through influence-aware text localization. QuISE then replaces these regions with two semantically distinct replacement texts that are irrelevant to both the query and the image. The final answer is determined by answer consistency across the edited images. Extensive experiments on three typographic-attack benchmarks, four attack settings, and four VLMs show that QuISE consistently improves defended accuracy. QuISE achieves a recovery rate of 67.9-75.0% with a harm rate of 0.5-1.1%.

One Modality to Forget Them All: Enhancing Cross-Modal Unlearning in Vision-Language Models

Machine unlearning is widely used to remove hazardous knowledge from large language models. Modern Vision-Language Models (VLMs), however, process both text and visual inputs, raising a fundamental security question: does unlearning in one modality transfer to the other? We present the first systematic, bidirectional study of cross-modal unlearning transfer across three VLM architectures: LLaVA-1.5 (MLP projection), InstructBLIP (Q-Former), and IDEFICS (gated cross-attention). We find that unlearning transfers across modalities, but the transfer is asymmetric and incomplete. In some cases, text unlearning strongly transfers to vision. However, this robustness is not preserved under typographic attacks that manipulate the visual presentation of text. Under such attacks, previously unlearned knowledge can be readily recovered, indicating shallow unlearning. To address the transfer gap and shallow robustness, we propose \textsc{CrossInf}, an influence-guided mitigation strategy. Motivated by the observation that different model components contribute unequally to cross-modal transfer, \textsc{CrossInf} focuses unlearning on transformer blocks that most influence cross-modal generalization. It reduces the transfer gap by more than half in architectures with strong fusion, while preserving model utility. It also improves robustness under typographic attacks, reducing the attack success rate to near zero. We further conduct human evaluation with three annotators ($κ{=}0.77$) to validate our findings. Finally, we analyze shallow unlearning using Centered Kernel Alignment (CKA), providing insights into the observed transfer behavior and robustness limitations.

What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks

Large language model (LLM)-powered content moderation systems have become a critical defense against harmful online content. However, these systems primarily operate on tokenized text and largely ignore the visual cues that humans naturally rely on when interpreting content. We show that this discrepancy creates a fundamental perceptual mismatch: content that is readily recognized as harmful by humans can become effectively invisible to automated moderation systems. To study this vulnerability, we introduce a class of Human-Perceptible Adversarial Attacks (HPAA), in which harmful expressions are embedded into otherwise benign text through visually salient typographic manipulations. Our key insight is that typographic features, including spacing, visual emphasis, and spatial arrangement, can be strategically combined to preserve human recognition of harmful content while substantially reducing machine detectability. Operating in black-box settings with only a small query budget, our attack automatically generates evasive content without requiring model access or gradient information. We evaluate the attack across multiple datasets and ten deployed moderation systems, including commercial APIs and state-of-the-art open-source guardrails. Results reveal a striking gap between human and machine perception: with only three detector queries, generated attacks achieve over 86\% human recognition while maintaining detection rates below 1\% across the evaluated systems. We further conduct ablation studies to identify the typographic factors driving successful evasion, analyze why current moderation architectures fail to capture these signals, and discuss practical defenses. Our findings expose a fundamental blind spot in today's LLM-based moderation ecosystem and highlight need for moderation systems that reason about content in a manner more consistent with human perceptual understanding.