Engineered as my Master’s Dissertation in Cybersecurity at the West University of Timișoara, this research evaluates state-of-the-art Deep Learning generative architectures for digital forensics, tamper restoration, deepfake detection, and synthetic forensic dataset generation.

Evaluated Architectures & Models

The research benchmarked four primary generative paradigms using PyTorch and CUDA acceleration on high-resolution datasets (FFHQ, LAION-5B, and ImageNet):

• StyleGAN3: Alias-free generative adversarial synthesis optimized for high-fidelity facial reconstruction ($FID = 2.89$, $PSNR = 28.1\text{ dB}$).
• Stable Diffusion: Latent diffusion denoising algorithms evaluated for contextually guided forensic image inpainting and restoration.
• VQ-VAE-2: Vector-quantized hierarchical variational autoencoders for discrete latent space representation and anomaly detection.
• DALL-E: Transformer-based autoregressive models evaluated for text-to-image synthesis and contextual accuracy.

Quantitative Metrics & Empirical Analysis

Algorithms were benchmarked across standardized computer vision metrics: Fréchet Inception Distance (FID), Inception Score (IS), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR) to determine trade-offs between mathematical reconstruction accuracy, generative fidelity, and computational overhead.