LiteUpdate: A Lightweight Framework for Updating AI-Generated Image Detectors

TL;DR

The paper tackles the problem of AI-generated image detectors degrading as generative models rapidly evolve. It introduces LiteUpdate, a lightweight framework consisting of a representative sample selection module and a model merging module to update detectors using limited new-data samples. The approach achieves improved detection performance and data efficiency across multiple detectors and generative models, while mitigating catastrophic forgetting of prior knowledge; notable results include substantial gains on AIDE with Midjourney (from 87.63% to 93.03%). The proposed method demonstrates strong practical potential for real-world deployment, enabling robust, scalable updates as generators continue to evolve.

Abstract

The rapid progress of generative AI has led to the emergence of new generative models, while existing detection methods struggle to keep pace, resulting in significant degradation in the detection performance. This highlights the urgent need for continuously updating AI-generated image detectors to adapt to new generators. To overcome low efficiency and catastrophic forgetting in detector updates, we propose LiteUpdate, a lightweight framework for updating AI-generated image detectors. LiteUpdate employs a representative sample selection module that leverages image confidence and gradient-based discriminative features to precisely select boundary samples. This approach improves learning and detection accuracy on new distributions with limited generated images, significantly enhancing detector update efficiency. Additionally, LiteUpdate incorporates a model merging module that fuses weights from multiple fine-tuning trajectories, including pre-trained, representative, and random updates. This balances the adaptability to new generators and mitigates the catastrophic forgetting of prior knowledge. Experiments demonstrate that LiteUpdate substantially boosts detection performance in various detectors. Specifically, on AIDE, the average detection accuracy on Midjourney improved from 87.63% to 93.03%, a 6.16% relative increase.

Figures (5)

• Figure 1: All traditional detectors pre-trained on SD 2.1 suffer a significant performance drop when applied to images from unseen generative models. Each subplot represents the detection performance of each AI-generated image detector evaluated on five different generators.
• Figure 2: The overview of LiteUpdate. It consists the representative sample selection module and the model merge module. In each step, $D_{\theta_0}$ is the pre-trained detector on generator $G$, while $D_{\theta_1}$ and $D_{\theta_2}$ is the fine-tuned detector on the genrator $G_{new}$ with the random and representative selection module. $D_{new}$ is the updated detector with the LiteUpdate framework.
• Figure 3: Examples of real-world and AI-generated images from our dataset. AI-generated images are from generators such as SD 2.1, SD 3.5, FLUX, Midjourney, and DALL-E 3 with the prompt "Carrot Cake Pancakes topped with Maple Cream Cheese drizzle". We can see that all samples exhibit similar quality in terms of content preservation.
• Figure 4: Efficiency comparison between random samples fine-tuning and the LiteUpdate on AIDE detector while the newly generator is SD 3.0. The LiteUpdate achieves superior detection performance using only 1,000 samples, surpassing the detector fine-tuned with random 5,000 samples.
• Figure 5: The mean detection accuracy of different strategies on AIDE across both old and newly generative models under a multi-generator setting. LiteUpdate refers to our framework designed to adapt to a single newly generator, while LiteUpdate-Multiple extends this capability to simultaneously adapt to multiple emerging generators.