Abstract:
In this paper, i present a study focused on effectively classifying images into two
categories: real-world images and artificially generated images using deep learning
techniques. To achieve this, i first gathered a public dataset named CIFAKE, which
comprises both real images captured from the real world and synthetic images generated
by artificial intelligence algorithms. I then developed a convolutional neural network
(CNN) with a custom attention module tailored to enhance classification accuracy. Our
approach involved comparing the performance of our custom CNN model with several
state-of-the-art CNN models commonly used for image classification tasks. These models
include VGG16, ResNet50, InceptionV3, MobileNet, and DenseNet. By evaluating these
models on the CIFAKE dataset, I aimed to discern the effectiveness of our proposed
architecture in distinguishing between real and AI-generated images. Furthermore, I
conducted a thorough analysis of the results using various statistical measurements to
assess the classification performance of each model. This analysis included metrics such
as accuracy, precision, recall, and F1 score. Through these statistical analyses, I sought to
provide insights into the strengths and limitations of different deep learning models for
image classification tasks, particularly in distinguishing between real-world and AIgenerated images. Overall, our study contributes to the advancement of image
classification techniques by addressing the increasingly relevant challenge of
differentiating between authentic real-world images and synthetic images generated by
artificial intelligence. The findings of this research could have significant implications for
various applications, including image forensics, content moderation, and computer vision
systems deployed in real-world scenarios.
