Robust Fake Image Detection via Incremental Training of Siamese Neural Networks

Aniello Castiglione, Lucia Cimmino, Michele Nappi, Matteo Polsinelli, Carlo Sorrentino

Abstract: The rapid advancement of generative models, particularly Generative Adversarial Networks (GANs) and Diffusion Models (DMs), has enabled the creation of highly realistic synthetic images that are increasingly difficult to distinguish from authentic ones. This progress presents significant challenges for digital content authentication. Furthermore, deepfake technologies are being actively exploited in cybersecurity attacks, introducing serious risks to information security. The use of hyper-realistic synthetic media in social engineering dramatically enhances the effectiveness of attacks such as identity theft and ransomware. Existing Deep Learning (DL)-based detectors often struggle to generalize to previously unseen generative techniques. To overcome this limitation, we propose a novel detection framework based on Siamese Neural Networks (SNNs), which focus on learning the similarity between image pairs rather than relying on fixed class boundaries. The proposed architecture is coupled with an incremental training strategy, where generators that yield the lowest detection performance are progressively incorporated into the training process. This enables the model to adapt more effectively to a diverse and evolving range of synthetic content. Experimental results show that strong generalization can be achieved using training data from as few as two generators, demonstrating the efficiency and scalability of the proposed approach. However, it is also observed that increasing the number of generators beyond a certain point does not lead to further improvements in performance. This indicates potential limitations in the current SNN training process. Future research will focus on refining training strategies to further improve generalization capabilities.

External IDs:dblp:conf/css/CastiglioneCNPS25