| Titre : |
Detection of AI-generated images using a lightweight CNN for resource-constrained environments |
| Type de document : |
document multimédia |
| Auteurs : |
Insaf Allaoui, Auteur ; Mustapha Bouakkaz, Directeur de thèse |
| Editeur : |
Laghouat : Université Amar Telidji - Département d'informatique |
| Année de publication : |
2026 |
| Importance : |
59 p. |
| Accompagnement : |
1 disque optique numérique (CD-ROM) |
| Note générale : |
Option : Artificial intelligence and data science |
| Langues : |
Anglais (eng) |
| Mots-clés : |
AI-generated image detection Lightweight deep learning Diffusion models Depthwise separable convolution Digital forensics Edge computing |
| Résumé : |
Given that models like Midjourney and Stable Diffusion are able to create images that are nearly impossible to distinguish from photographs, deep learning models like ResNet and VGG can be used to detect these fakes. However, the computational requirements of these models make them impractical for use on resource-limited devices. To address this, a lightweight convolutional neural network can be developed for the purpose of AI image detection.
By utilizing Depthwise Separable Convolutions and Global Average Pooling, the model was able to drastically reduce the number of parameters that the model had to learn in order to correctly identify the images created by AI art generators. Furthermore, the model was trained from scratch using a dataset composed of both authentic photographs and synthetic AI art images, and was additionally benchmarked against models like ResNet50, VGG16, and DenseNet121.
Results indicated that the proposed model was able to achieve an accuracy of 84.33% with only 649,985 parameters, which is 11 times fewer parameters than DenseNet121 and 36 times fewer parameters than ResNet50. Additionally, with an average inference time of 14 milliseconds, this model indicates that it may be viable for deployment in real-time applications on resource-limited devices. |
| note de thèses : |
Mémoire de master en informatique |
Detection of AI-generated images using a lightweight CNN for resource-constrained environments [document multimédia] / Insaf Allaoui, Auteur ; Mustapha Bouakkaz, Directeur de thèse . - Laghouat : Université Amar Telidji - Département d'informatique, 2026 . - 59 p. + 1 disque optique numérique (CD-ROM). Option : Artificial intelligence and data science Langues : Anglais ( eng)
| Mots-clés : |
AI-generated image detection Lightweight deep learning Diffusion models Depthwise separable convolution Digital forensics Edge computing |
| Résumé : |
Given that models like Midjourney and Stable Diffusion are able to create images that are nearly impossible to distinguish from photographs, deep learning models like ResNet and VGG can be used to detect these fakes. However, the computational requirements of these models make them impractical for use on resource-limited devices. To address this, a lightweight convolutional neural network can be developed for the purpose of AI image detection.
By utilizing Depthwise Separable Convolutions and Global Average Pooling, the model was able to drastically reduce the number of parameters that the model had to learn in order to correctly identify the images created by AI art generators. Furthermore, the model was trained from scratch using a dataset composed of both authentic photographs and synthetic AI art images, and was additionally benchmarked against models like ResNet50, VGG16, and DenseNet121.
Results indicated that the proposed model was able to achieve an accuracy of 84.33% with only 649,985 parameters, which is 11 times fewer parameters than DenseNet121 and 36 times fewer parameters than ResNet50. Additionally, with an average inference time of 14 milliseconds, this model indicates that it may be viable for deployment in real-time applications on resource-limited devices. |
| note de thèses : |
Mémoire de master en informatique |
|  |
Detection of AI-generated images using a lightweight CNN for resource-constrained environments
