Integrating Advanced Feature Extraction and LIME for Real and AI Generated Image Classification

Advancement of deep network generative models has led to producing highly quality realistic images a reality. This also has started to create fake content which may be used for illegal and antisocial purposes. Therefore, its important to have tools which can help determine the authenticity of the images available to the consumers. This work addresses this highly relevant problem by employing spectral features such as Discrete Cosine Transform (DCT), Fast Fourier Transform (FFT), and Wavelet Transform, along with statistical classifiers in differentiating the real images from the fake ones. The results of this approach is compared against two advanced convolutional neural network (CNN) architectures, namely, ResNet-50 & EfficientNet-B0. Incorporating Local Interpretable Model-Agnostic Explanations (LIME) provides a visible insight for the model predictions emphasizing on various localized features and image areas contributing significantly to the classification process. ResNet-50 emerged as the best performer with an accuracy of 94.7% on the test dataset followed by EfficientNet-B0 at 94.1%. Among the statistical classifiers, the CatBoost emerged as the leader with an accuracy of 83.71%.