Evolution of Detection Performance throughout the Online Lifespan of Synthetic Images

TL;DR

This work tackles the problem of detecting AI-generated synthetic images that spread online, revealing that state-of-the-art detectors struggle in real-world conditions and deteriorate as content is repeatedly post-processed over time. It introduces the FOSID dataset to capture web-scale evolution of online misinformation and systematically evaluates a broad set of SID methods, exposing calibration gaps and degradation in the wild. To mitigate this, the authors propose Retrieval-Assisted Synthetic Image Detection (RASID), which leverages near-duplicate image retrieval to stabilize detection across the image lifespan, achieving average gains of $6.7 ext{ extpercent}$ in $BA$ and $7.8 ext{ extpercent}$ in $AUC$. The results highlight the need for evolution-aware, retrieval-informed SID approaches and provide a publicly available benchmark to drive future improvements in misinformation defense against diffusion-model–generated imagery.

Abstract

Synthetic images disseminated online significantly differ from those used during the training and evaluation of the state-of-the-art detectors. In this work, we analyze the performance of synthetic image detectors as deceptive synthetic images evolve throughout their online lifespan. Our study reveals that, despite advancements in the field, current state-of-the-art detectors struggle to distinguish between synthetic and real images in the wild. Moreover, we show that the time elapsed since the initial online appearance of a synthetic image negatively affects the performance of most detectors. Ultimately, by employing a retrieval-assisted detection approach, we demonstrate the feasibility to maintain initial detection performance throughout the whole online lifespan of an image and enhance the average detection efficacy across several state-of-the-art detectors by 6.7% and 7.8% for balanced accuracy and AUC metrics, respectively.

Figures (3)

• Figure 1: We study the evolution of synthetic images throughout their online lifespan through collecting different online versions of the "same" synthetic image. Using this data, we evaluate state-of-the-art synthetic image detectors, to find out that they mostly fail to detect several instances that were shared online, while the time since initial sharing negatively affects detection performance. Using retrieval-assisted synthetic image detection, it is feasible to retain the initial detection performance throughout the online lifespan of a synthetic image.
• Figure 2: Fact-checked synthetic images used as seeds for the data collection process. The "Pope" image is a satirical depiction of the Pope, while the remaining three were presented as relating to events of the Israel-Hamas war. The images have been cropped and scaled to the same aspect ratio for illustration purposes in this figure only.
• Figure 3: Examples of basic and non-basic images from the Gaza1 subset of FOSID.