Leveraging Hierarchical Image-Text Misalignment for Universal Fake Image Detection

TL;DR

The paper tackles the lack of generalization in fake image detectors by reframing detection as a multimodal problem. It introduces ITEM, which measures image-text misalignment in the CLIP joint space and augments this with a hierarchical global-plus-local scheme to capture both broad and fine-grained semantic clues. An MLP head is trained to classify real versus fake based on the misalignment distance, with robust results showing improved generalization to unseen models and resilience to perturbations. The approach demonstrates that leveraging vision-language representations yields stronger universal detection capabilities, offering practical benefits for safeguarding against AI-generated content.

Abstract

With the rapid development of generative models, detecting generated fake images to prevent their malicious use has become a critical issue recently. Existing methods frame this challenge as a naive binary image classification task. However, such methods focus only on visual clues, yielding trained detectors susceptible to overfitting specific image patterns and incapable of generalizing to unseen models. In this paper, we address this issue from a multi-modal perspective and find that fake images cannot be properly aligned with corresponding captions compared to real images. Upon this observation, we propose a simple yet effective detector termed ITEM by leveraging the image-text misalignment in a joint visual-language space as discriminative clues. Specifically, we first measure the misalignment of the images and captions in pre-trained CLIP's space, and then tune a MLP head to perform the usual detection task. Furthermore, we propose a hierarchical misalignment scheme that first focuses on the whole image and then each semantic object described in the caption, which can explore both global and fine-grained local semantic misalignment as clues. Extensive experiments demonstrate the superiority of our method against other state-of-the-art competitors with impressive generalization and robustness on various recent generative models.

Figures (9)

• Figure 1: Motivation behind our method. We find that the generated fake images cannot properly align with corresponding captions compared to real images, which could serve as clues for a more general and robust universal detector.
• Figure 2: Overview of our proposed method. We explore the misalignment between image and text modalities on both the global semantic clues, i.e., the whole image, full caption, and local fine-grained semantic clues, i.e., each local semantic object. After the representation learning stage, we optimize an MLP head to perform the usual fake-image detection task.
• Figure 3: Image-text misalignment representation. High cosine similarity may lead to significant distance in original CLIP space (a). Our defined misalignment representation $\mathbf{D}$ is less significant under high cosine similarity and more significant under low cosine similarity, which could properly respond to the modality misalignment.
• Figure 4: Robustness results to unseen perturbations. Average precision (AP) under three different types of perturbations with three different severity levels: Gaussian Noise ($\sigma = 0.001,0.005,0.01$), Gaussian Blur ($\sigma = 1,2,3$), and JPEG Compression ($quality = 75,50,25$) (from left to right).
• Figure 5: Ablation study on different distances. The average precision (AP(%)) is reported. We observe that our proposed local and global distances could both achieve impressive performance, and the performance is further boosted with equipped both.