A Novel Approach for Differentiating AI-Generated and Real Images Using Gabor and Color Layout Filters (GF and CLF)

Abstract

The emergence of generative AI images has profoundly upended the art world. Distinguishing AI-generated images from human-created artwork is increasingly challenging. If this issue remains unresolved, dishonest individuals may exploit those who are willing to pay more for original artwork and businesses whose stated policies exclude AI graphics. There are a number of methods for differentiating AI photos from human art: Diffusion model-focused research tools, classifiers developed through supervised learning, and identification by experienced artists utilizing their understanding of creative methods. The objective of this study is to evaluate the performance of various machine learning algorithms when applied to datasets enhanced through two image enhancement techniques: Gabor Filter (GF) and Colour Layout Filter (GF). The focus is on comparing the effectiveness of these filters in improving the accuracy, precision, recall, ROC, and PRC of selected algorithms, thereby determining which technique yields superior results for different machine learning models. The problem statement addresses the challenge of optimizing machine learning model performance on image datasets. Specifically, it investigates whether the Gabor Filter, known for its effectiveness in feature extraction from images, can outperform the CLF Filter in enhancing the predictive capabilities of algorithms such as Bayes Net, Sequential Minimal Optimization (SMO), Instance-Based K-nearest neighbor (IBK), Bagging, Jrip, and Random Forest. The filters assist in retrieving more pertinent features in the dataset that subsequently increases both the model robustness and the accuracy of classification. Several performance measures were relied upon to evaluate the models and determine their comparative performance: The accuracy, the precision, the recall, the ROC area (Receiver Operating Characteristic), and the PRC region (Precision-Recall Curve). The best model was Random Forest with Gabor Filter (RF-GF) that got the highest accuracy (92.43), precision score (0.94), recall score (0.93), ROC (0.96), as well as PRC (0.99). The tests that provide statistically significant responses were Pairwise Wilcoxon signed-rank tests indicating that RF-GF performed significantly better when compared to other models, including SMO with Gabor Filter (SMO-GF) and CLF-based models. The proposed model was also compared to state-of-the-art methods and external benchmarks by the study proving the competitiveness of the model in regard to performance and efficiency of calculations. Moreover, a combination of the application in the AI detecting system like image classification, fraud detection, and medical diagnosis was considered. The findings show that the RF-GF model is resilient, effective and performant that can be used in real-life applications whereby there are constraints to the available computational resources.