AI-based image generation: The impact of fine-tuning on fake image detection

Abstract: Machine learning-based image generation models such as Stable Diffusion are now capable of generating synthetic images that are difficult to distinguish from real images, which gives rise to a number of legal and ethical concerns. As a potential measure of mitigation, it is possible to train neural networks to detect the digital artifacts present in the images synthesized by many generative models. However, as the artifacts in question are often rather model-specific, these so-called detectors usually suffer from poor performance when presented with images from models it has not been trained on. In this thesis we study DreamBooth and LoRA, two recently emerged finetuning methods, and their impact on the performance of fake image detectors. DreamBooth and LoRA can be used to fine-tune a Stable Diffusion foundation model, which has the effect of creating an altered version of the base model. The ease with which this can be done has led to a proliferation of communitygenerated synthetic images. However, the effect of model fine-tuning on the detectability of images has not yet been studied in a scientific context. We therefore formulate the following research question: Does fine-tuning a Stable Diffusion base model using DreamBooth or LoRA affect the performance metrics of detectors trained on only base model images? We employ an experimental approach, using the pretrained VGG16 architecture for binary classification as detector. We train the detector on real images from the ImageNet dataset together with images synthesized by three different Stable Diffusion foundation models, resulting in three trained detectors. We then test their performance on images generated by fine-tuned versions of these models. We find that the accuracy of detectors when tested on images generated using fine-tuned models is lower than when tested on images generated by the base models on which they were trained. Within the former category, DreamBooth-generated images have a greater negative impact on detector accuracy than LoRA-generated images. Our study suggests there is a need to consider in particular DreamBooth fine-tuned models as distinct entities in the context of fake image detector training.