Fine-tuning convolutional neural network for artificial intelligence generated image detection enhancement

Abstract

The relationship between art and technology has changed how people engage with creativity, leading to the industrialization of the field. Various digital media have been utilized in the endeavour of art creation, such as artificial intelligence (AI) generation for images. The utilization of AI-generated art has yielded negative reactions due to its exploitative nature on pre-existing artworks without the creator’s consent, which raises plagiarism concerns. This research utilized convolutional neural network (CNN) to help detect such images to reduce public concerns on the abuse of AI images. The algorithm is proposed to detect such images as it involves spatial convolution within two-dimensional spaces, matching the nature of images. The model was developed from pre-existing architectures, namely EfficientNetB1 and Xception, which was pre-trained on ImageNet classification task with the modification of inclusion or exclusion of dropout in the top layer. After assessing the models, removing top layer dropout from EfficientNetB1 model improved it to reach the F1-score of 97.66% compared to 97.44% in the base model and Xception with a dropout layer yields lower F1-score of 95.56% compared to 97.07% in the base model.