This study presents a hierarchical multi-level approach for deepfake detection and recognition, achieving over 97% classification accuracy. The approach distinguishes between images generated by Generative Adversarial Networks (GANs) and Diffusion Models (DMs), and further identifies the specific architectures used to generate synthetic data. A dataset of 83,000 images, including 9 GAN architectures and 4 DMs, was used for training. The hierarchical framework consists of three levels: Level 1 classifies real vs. AI-generated images, Level 2 distinguishes between GANs and DMs, and Level 3 identifies specific architectures. The models achieved high accuracy across all levels, with Level 3-DM achieving 99.37% accuracy and Level 3-GAN achieving 97.01%. The approach is robust to various attacks, including JPEG compression, resizing, and rotation, and can be applied in real-world contexts such as forensic investigations. The hierarchical method outperforms flat classification models by improving accuracy by about 2%. The study also demonstrates the method's effectiveness on datasets like COCOfake and FaceForensics++, showing its potential for enhancing deepfake detection systems and safeguarding digital media authenticity.This study presents a hierarchical multi-level approach for deepfake detection and recognition, achieving over 97% classification accuracy. The approach distinguishes between images generated by Generative Adversarial Networks (GANs) and Diffusion Models (DMs), and further identifies the specific architectures used to generate synthetic data. A dataset of 83,000 images, including 9 GAN architectures and 4 DMs, was used for training. The hierarchical framework consists of three levels: Level 1 classifies real vs. AI-generated images, Level 2 distinguishes between GANs and DMs, and Level 3 identifies specific architectures. The models achieved high accuracy across all levels, with Level 3-DM achieving 99.37% accuracy and Level 3-GAN achieving 97.01%. The approach is robust to various attacks, including JPEG compression, resizing, and rotation, and can be applied in real-world contexts such as forensic investigations. The hierarchical method outperforms flat classification models by improving accuracy by about 2%. The study also demonstrates the method's effectiveness on datasets like COCOfake and FaceForensics++, showing its potential for enhancing deepfake detection systems and safeguarding digital media authenticity.