PetalGuard - Private Federated Learning.

TII provides expert services in deploying confidential computing solutions for secure federated learning. Our proprietary framework, PetalGuard, integrates federated learning with secure Multi-Party Computation (MPC) to ensure end-to-end data confidentiality. PetalGuard enables decentralized model aggregation without exposing training data or model parameters of the clients. Participants train models locally; neither raw data nor model updates leave their environment unprotected. All model parameters are secret-shared before transmission, and aggregation is performed entirely in secret-shared form, ensuring continuous protection of sensitive information.

Our services include end-to-end deployment of PetalGuard on client premises or in private cloud environments, tailored configuration to meet specific use case requirements, and comprehensive support throughout the federated learning process.

Training AI models requires access to large, often fragmented datasets owned by multiple entities and governed by varying privacy regulations. Decentralized methods like federated learning offer partial data protection, but they still leave room for privacy risks, especially the possibility of inferring sensitive information from locally trained models. PetalGuard addresses this challenge by integrating Multi-Party Computation, enabling secure, privacy-preserving model training. It represents a significant advancement in safeguarding data during collaborative AI development.

PetalGuard leverages AWS’s robust global infrastructure to deliver 24/7 access to secure and resilient data center capacity. Building across the more than 114 availability zones around the world, allows the platform to scale from hundreds to thousands of servers. PetalGuard utilizes AWS compute, including EC2 and EKS, to provide access to accelerated virtual machines and containers. With GPU-enabled EC2 instances, PetalGuard can take advantage of the AWS infrastructure to run training on-demand, and scale down to zero when training isn't required.