Deepfake Image Detection Using GWO-Based Hypertuning and ResNet-RS Architecture-Based Classification

Abstract

Online platforms are greatly endangered by deepfakes, which are synthetic pictures created using deep generative models. This has prompted a number of research initiatives aimed at improving deepfake picture detection, with promising results on publically accessible datasets. Our analysis of eight cutting-edge detectors leads us to believe that, as a result of two new advancements, they are still not prepared for deployment. First, the possibility of an attacker creating several customized generators (to make deepfakes) has the potential to significantly increase the danger surface, especially with the advent of lightweight tools to alter big generative models. We prove that current safeguards aren’t up to the task of protecting publicly accessible user-customized generative models. We go over some novel ML techniques that leverage ensemble modeling and content-agnostic features to outperform user-customized models in generalization. Furthermore, hostile deepfakes that circumvent current defenses may be created by exploiting the proliferation of vision foundation models, which are models based on machine learning trained on generalized data that can be readily adjusted to do several downstream tasks. By meticulously manipulating the semantic content of the images, we provide a straightforward adversarial strategy that uses pre-existing foundation models to create adversarial samples free of adversarial noise. We point out where various defenses are weak in the face of our assault and discuss potential ways forward that make use of adversarial training and sophisticated foundation models.

Keywords: Deepfake images, GAN model, Generative models, GWO tuning.

View PDF