Generalisering och robusthet vid detektering av syntetiska bilder med en enda enkel del av bilden
Year: 2025
Type: dissertation
Abstract: Machine learning models have become increasingly powerful tools for image classification, including tasks such as identifying manipulated or AI-generated content. As generative AI systems advance, especially through methods like GANs and diffusion models, they enable the creation of hyper-realistic synthetic images, posing new challenges for media authentication and raising concerns about digital misinformation. Reliable detection systems are therefore critical to preserving the integrity of digital media. Traditional deepfake detection methods often rely on full-image classification, which can be ineffective when generative models introduce only subtle artifacts. Patch-based detection methods offer a promising alternative by focusing on localized image regions that may better reveal synthetic inconsistencies. This study investigates whether incorporating the Single Simple Patch (SSP) method improves the accuracy and generalization of machine learning models in detecting AI-generated images. Two architectures, a simple CNN and a pre-trained ResNet50, were evaluated on the GenImage dataset, which includes images from several generative models. Each model was tested both with and without the SSP method. Results showed that SSP significantly increased detection accuracy when used with ResNet50, while the simple CNN failed to benefit meaningfully. Among the generators, Midjourney images were the hardest to detect when training on other sources, while GLIDE images were the easiest. The study also assessed the robustness of the ResNet50 model with SSP by applying common image edits such as blurring, JPEG compression, and additive noise. Even low levels of corruption led to a sharp drop in accuracy, confirming that SSP is highly sensitive to image distortions. While SSP enhances detection on clean images, its vulnerability to simple edits limits its practical deployment in uncontrolled environments.
Authors: Alvarez Ohlin, Oscar, Hellvig, Jonatan
Institution: KTH Royal Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7078796477
