A hybrid approach for the detection of images generated with multi generator MS-DCGAN

Sürücü S., DİRİ B.

Engineering Science and Technology, an International Journal, cilt.63, 2025 (SCI-Expanded) identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 63
  • Basım Tarihi: 2025
  • Doi Numarası: 10.1016/j.jestch.2025.101969
  • Dergi Adı: Engineering Science and Technology, an International Journal
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, INSPEC, Directory of Open Access Journals
  • Anahtar Kelimeler: Deep convolutional generative adversarial network, Generative adversarial networks, Manipulation detection, Multi-spectral image, Multi-spectral image generation, Remote sensing
  • Yıldız Teknik Üniversitesi Adresli: Evet

Özet

Over the past few years, there have been significant advances in remote sensing technology that have considerably expanded the range of research that can be conducted using remote sensing systems. Various fields, from agriculture to defense applications, use remote sensing imagery, primarily acquired by sensors mounted on vehicles like satellites and UAVs. In addition to advances in remote sensing technology, there have also been major advancements in deep learning. In recent years, there has been a substantial increase in the studies on these two topics. Generative Adversarial Networks (GAN) technology, another area of artificial intelligence and deep learning research, has taken the generation of fake satellite images to a new level. Users can use these artificial images for a variety of purposes, including information concealment and data expansion. Malicious uses of the generated fake images could trigger international crises. In this paper, we propose a new method for the generation and detection of fake satellite images. The MultiSpectral Deep Convolutional GAN (MS-DCGAN) model is developed to generate fake multispectral images, and the TransStacking model is proposed to distinguish between fake images and real images. This model is tested both as a single generator and multi generator model. The TransStacking (DenseNet201+stacking) model showed a very high success rate achieving 100% accuracy for single generator and 98% accuracy for multi generator MS-DCGAN, respectively. The proposed model is an advanced hybrid model that provides the best results in multi-spectral images and can be applied in diverse domains. Since the TransStacking model is a modular hybrid model, it can be used with many different old and new models. Furthermore, the effect of the models in the base part of the stacking module on the results was also analyzed by performing ablation analysis on the DenseNet201+stacking model, where the best results were obtained.