Beyond Pixels: Semantics-Aware Typographic Attacks for Geographic Privacy Protection

The researchers propose a semantics-aware typographic attack that adds deceptive text outside the image frame to interfere with large vision-language models (LVLMs) inferring a user's geographic location from social media images, thereby protecting geographic privacy. The paper points out that LVLMs can infer a user's location directly from shared images, constituting an overlooked privacy threat; adversarial image perturbations, however, require strong distortion to be effective against LVLMs, which noticeably degrades image quality and reduces the image's sharing value. The method uses a two-stage design and investigates which text semantics can effectively disrupt localization inference. Experiments on three datasets show that the method significantly reduces the geographic location prediction accuracy of five state-of-the-art commercial LVLMs, forming a practical protection strategy that preserves image quality.