Analyzing and Improving Existing Neural Network-Based Approaches to Identify AI Generated Images

Abstract

Images produced by AI software, particularly human faces, have the potential to spread misinformation, thus detection tools are essential to tell facts from fiction. However, current neural network-based tools misclassify images with extreme photography conditions, facial obstructions, or post-processing. This project aimed to address the limitations of existing detection models, especially regarding poorly-photographed or edited images. A MobileNet-v2-based neural network to identify synthetic images was analyzed for possible shortcomings. After obtaining its baseline accuracy, it was stress tested on validation data with brightness, contrast, hue, and saturation edited by set increments. The model performed poorly on processed images and faces with glasses or shadows. Possible limitations were insufficient data augmentation, the Global Average Pooling Layer, too few epochs trained, and inadequate regularization strength or dropout rate. The model was then modified to address these shortcomings by altering the brightness, contrast, hue, and saturation of training data, adding Gaussian noise to training images, implementing a Global Max Pooling Layer, augmenting the number of epochs trained, and changing the regularization strength and dropout rate. The modified model was evaluated on augmented and non-augmented images. Overall, the model improved when Gaussian noise was added with standard deviation 0.1, it was trained for 15 epochs, and there were no dropout layers. This was because the Gaussian noise layer approximated degraded image quality, and the model was too simple to learn image features when trained for fewer epochs or when neurons were disabled. This research could be expanded by testing multiple modifications at once.