On the Use of Mixup Augmentation for Synthetic Image Attribution in the Wild

Due to the continuous proliferation of generative models and the often unauthorized dissemination of synthetic images, the problem of tracing back the origin of a synthetic image, namely, synthetic image attribution, has raised a lot of interest. Despite the recent advances in this field, performing synthetic image attribution in the wild is still challenging. Recent literature on Open Set Recognition (OSR) in machine learning has shown that training augmentation has a strong effect on the open-set performance of OSR methods. Inspired by this literature, in this paper, we assess the impact on the synthetic image attribution task of a new type of augmentation, called mixup augmentation, that consists of interpolating data from different classes. Our experiments with models based on Convolutional Neural Networks (CNNs) as well as Vision Transformers (ViTs), reveal that this type of augmentation is indeed effective in improving robustness and generalization of synthetic image attribution classifiers, yet at the expense of a reduction of open-set performance. We also proposed a new variant of this type of augmentation, performing mixup of the samples with frozen labels, that keeps the benefits of the standard mixup in terms of robustness and generalization, at the same time mitigating the performance drop in open-set.

I. Introduction

The rapid diffusion and progress of generative models, such as diffusion models (DMs) and generative adversarial networks (GANs) has led to the widespread use of generated images. This has caused the unauthorized dissemination of synthetic images across social networks. In response to the threats posed by forged images, significant research efforts have been made in the field of detecting forgeries [1]. Besides determining whether an image is real or fake (synthetic image detection), understanding the provenance (origin) of an image, referred to as synthetic image attribution, also plays an important role. Several methods have been proposed for model-level attribution, that rely on the artefacts or signatures (fingerprints) left by the models in the images they generate [2]. Recent works have started addressing the attribution task in a different manner, to attribute the synthetic images to the source architecture that generated them, instead of the specific model [3]. Such an approach overcomes a limitation of model attribution in real-world applications, where model-level granularity is often not needed and even undesired (e.g. if an attacker steals a copyrighted GAN, and modifies the weights by fine-tuning on a different dataset, model-level attribution is going to fail).