ADVANCED DEEP LEARNING TO GENERATE AND DETECT FAKE IMAGES OF EGYPTIAN MONUMENTS

Abstract

This study examined the use of StyleGAN to create synthetic images of Egyptian monuments, addressing a critical gap at the intersection of generative artificial intelligence and cultural heritage. Through extensive experiments on a large dataset containing 5,000 Egyptian monument images, we show that architectural changes to the StyleGAN framework can significantly improve the quality and authenticity of the generated images. Our study contributes to the existing literature. First, we designed an enhanced discriminator architecture incorporating noise injection, squeeze-and-excitation blocks, and an improved MinibatchStdLayer, resulting in a Fréchet Inception Distance 27.5% better than that of the original model. We further introduced a novel image-text alignment approach using SigLIP, which can generate semantically guided monuments. We applied Differential Evolution (DE) to optimize the latent space of the conditional generator to reduce the alignment error by 15% for the targeted monument-generation tasks. We systematically analyzed various truncation methods used to manage noise in generated images by finding the best parameters that fit the architecture best but are also diverse. Statistical validation using bootstrap confidence intervals, McNemar’s test and DeLong’s ROC analysis show significant improvements with effect sizes in the moderate to large range (Cohen’s d ≈ 0.9-1.4) The discriminator was able to achieve 95.5% accuracy with a 5.3% false positive rate and 3.6% false negative rate. This 62% error drop was compared to the baseline. Under heavily corrupted conditions (JPEG quality = 10; Gaussian blur σ = 5.0), it achieved 78-85% of the baseline performance, whereas the default achieved 65-72% of the baseline performance. Frequency domain analysis results revealed resilience, with AUC values generally >0.95, varying by frequency. The new discriminator was approximately 20 to 25 percent more robust to adversarial attacks. However, both architectures are fundamentally vulnerable to stronger attacks. Our research shows how strategic refinements of operations models can produce representations of Egyptian monuments that attain a high-quality and satisfactory level of diversity that we can detect. Innovations can greatly help in the preservation of cultural heritage, virtual tourism, visualization, and education. This study will allow the generation of high-quality and varied Egyptian monument images, which can help in the digital conservation and easy accessibility of one of the world’s great architectural heritages.