A Siamese-based Verification System for Open-set Architecture Attribution of Synthetic Images

TL;DR

The paper addresses open-set architecture attribution for synthetic images and proposes a Siamese verification framework using EfficientNet-B4 backbones to determine whether two images share the same generating architecture or to verify a claimed architecture against references. It introduces a two-phase training regime with contrastive embedding learning followed by a decision-head classifier, and evaluates on a diverse set of GAN, diffusion, and transformer architectures, focusing on synthetic faces. Key findings show strong open-set performance, robust generalization to unseen models, and superior results compared to open-set classifiers with rejection when the verifier is repurposed as a classifier. The work advances forensic capabilities by enabling reliable attribution in open-set scenarios without relying on a rejection option, with practical implications for security and integrity of digital media.

Abstract

Despite the wide variety of methods developed for synthetic image attribution, most of them can only attribute images generated by models or architectures included in the training set and do not work with unknown architectures, hindering their applicability in real-world scenarios. In this paper, we propose a verification framework that relies on a Siamese Network to address the problem of open-set attribution of synthetic images to the architecture that generated them. We consider two different settings. In the first setting, the system determines whether two images have been produced by the same generative architecture or not. In the second setting, the system verifies a claim about the architecture used to generate a synthetic image, utilizing one or multiple reference images generated by the claimed architecture. The main strength of the proposed system is its ability to operate in both closed and open-set scenarios so that the input images, either the query and reference images, can belong to the architectures considered during training or not. Experimental evaluations encompassing various generative architectures such as GANs, diffusion models, and transformers, focusing on synthetic face image generation, confirm the excellent performance of our method in both closed and open-set settings, as well as its strong generalization capabilities.

Figures (3)

• Figure 1: Verification scenarios considered in the paper.
• Figure 2: High-Level Architecture for the verification task.
• Figure 3: Examples of synthetic images from the 10 architectures.