Explore a cutting-edge deepfake detection showcase from CodeZero, built in partnership with AWS and using the Kaggle Deepfake Challenge dataset.
We deployed three distinct machine learning techniques on world-class AWS p4d.e24xlarge infrastructure to build models capable of identifying deepfake images.
ML Techniques Explored:
- Convolutional Neural Networks (CNNs): EfficientNet & ResNet, achieving our highest accuracy of 78.5%.
- Generative Adversarial Networks (GANs): Deep & Wasserstein GANs, providing a different approach to anomaly detection.
- Neural Radiance Fields (NeRF): A novel technique for generating 3D scenes to understand image authenticity.
All codebases utilized Distributed Data Parallel (DDP) to leverage the full power of the 8x NVIDIA A100 GPUs.
The AWS Architecture:
- Compute: AWS p4d instance (8x A100s, 96 vCPUs, 1TB RAM, 400Gbps network)
- Networking: Best-practice VPC in N. Virginia with CloudFront CDN.
- Storage: A tiered approach using NVMe, EBS, and EFS for optimal performance.
- Data Sync: S3 buckets for seamless data transfer.
- Management: SSM Fleet Manager, Bastion Jump Host, and CloudWatch for monitoring.
