Physics-Based Analysis and Data-Driven Detection of Artificial Intelligence–Generated Surface Crack Images for Civil Infrastructure Monitoring

Generative artificial intelligence (AI) is transforming the condition assessment of civil infrastructures, creating both opportunities and challenges. This paper aims to address challenges related to the authenticity and trustworthiness of AI-generated images within the context of surface cracking in civil infrastructure. Two state-of-the-art generative AI models—a denoising diffusion probabilistic model (DDPM) and style generative adversarial network–adaptive discriminator augmentation (StyleGAN3-ada)—are employed to synthesize realistic surface cracks commonly observed in concrete and asphalt pavement. Both models are trained using two data sets comprising red-green-blue (RGB) and masked versions of surface crack images. The quality of the generated images is quantitatively evaluated using the Fréchet inception distance (FID), with DDPM achieving lower FID scores (30.1 for RGB and 22.4 for masked) compared to the StyleGAN3-ada models (45.9 and 52.2, respectively). Given the superior performance of DDPM, evidenced by lower FID scores, the remainder of the analyses is conducted using images generated by the DDPM. The generated images are then analyzed in terms of their similarity to and detectability from real images. The analysis statistically compares the similarity of DDPM-generated images to real ones, focusing on physics-based critical features such as crack width and orientation. This paper uses the novel idea of crack-to-graph conversion to calculate the angular orientation of the crack images. Once the crack patterns are converted to their representative graphs, the angles of the graph edges are calculated, representing the cracks’ orientations. Findings show that the mean, median, and standard deviation of the width and orientation of the generated crack images have less than a 3% difference from that of the real images. To quantify the detectability of DDPM-generated images, two approaches are investigated: supervised convolutional neural network (CNN)-based classifiers to distinguish between real and DDPM-generated images and unsupervised autoencoders trained on real images, treating DDPM-generated images as anomalies. The detection models achieved precision, recall, and F1-scores exceeding 0.85 across both approaches, demonstrating their robustness. Although DDPM-generated surface cracks exhibit physical features similar to real images that make them visually indistinguishable by the human eye, they remain detectable through both supervised and unsupervised machine learning methods. Notably, the unsupervised approach offers the advantage of identifying AI-generated images without requiring them for training.