CaptionFool: Universal Image Captioning Model Attacks

TL;DR

CaptionFool introduces a universal adversarial attack on image captioning models that modifies a small portion of the image to generate targeted, potentially offensive captions, revealing significant vulnerabilities in current models.

Contribution

This paper presents the first universal, input-agnostic attack on transformer-based image captioning models, demonstrating high success rates with minimal image modifications.

Findings

Achieves 94-96% success rate in targeted caption generation

Modifies only about 1.2% of image patches for attack

Can generate offensive content to bypass moderation filters

Abstract

Image captioning models are encoder-decoder architectures trained on large-scale image-text datasets, making them susceptible to adversarial attacks. We present CaptionFool, a novel universal (input-agnostic) adversarial attack against state-of-the-art transformer-based captioning models. By modifying only 7 out of 577 image patches (approximately 1.2% of the image), our attack achieves 94-96% success rate in generating arbitrary target captions, including offensive content. We further demonstrate that CaptionFool can generate "slang" terms specifically designed to evade existing content moderation filters. Our findings expose critical vulnerabilities in deployed vision-language models and underscore the urgent need for robust defenses against such attacks. Warning: This paper contains model outputs which are offensive in nature.