What’s New in FLARE v2.7.0

The new features can be divided into following categories:

Confidential Federated AI

Confidential Federated AI Applications:

• Collaborative R&D & Analytics: Joint model training and agent-based analytics inside Confidential Computing enclaves, protecting data and IP.

• Model Protection: Secure use and fine-tuning of proprietary or licensed foundation models.

• Regulated & Cross-Border AI: Enable collaboration across finance, healthcare, defense, and global partners while maintaining privacy, sovereignty, and compliance.

• Secure Deployment: Prevent model or data leakage during inference in untrusted environments.

With this release, we offer this first-of-its-kind product for end-to-end IP protection solution in federated setup using confidential computing.

• The solution is for on-premise deployment on bare metal using AMD CPU and NVIDIA GPU with Confidential VM.

• End-to-End Protection: By end-to-end protection, we mean that we not only protect the IP (model and code) in use at runtime, but also protect against CVM tampering at deployment.

• The solution is able to perform

  • Secure aggregation on the server-side to protect against privacy leaks via model

  • Model theft protection on the client-side to safeguard Model IP during collaboration

  • Data leak prevention on the client-side with pre-approved, certified code.

Confidential Federated AI

This feature is in Technical Preview. Reach out to the NVIDIA FLARE team for CVM build scripts: federatedlearning@nvidia.com

You can read more about the user usage at FLARE Confidential Federated AI

FLARE Core

Job Recipe

Here is an example of the FedAvg Job Recipe

n_clients = args.n_clients
num_rounds = args.num_rounds
batch_size = args.batch_size

recipe = FedAvgRecipe(
    name="hello-pt",
    min_clients=n_clients,
    num_rounds=num_rounds,
    model=SimpleNetwork(),
    train_script="client.py",
    train_args=f"--batch_size {batch_size}",
)
add_experiment_tracking(recipe, tracking_type="tensorboard")

env = SimEnv(num_clients=n_clients)
run = recipe.execute(env)
print()
print("Result can be found in :", run.get_result())
print("Job Status is:", run.get_status())
print()

Introducing the new Flare Job Recipe: a lightweight way to capture the code needed to specify the client training logic and the server-side algorithm. The same Job Recipe can run seamlessly in SimEnv, PoCEnv, or ProdEnv—from local experiments to production deployments.

With Flare Job Recipe, we are making the federated learning workflow dramatically simpler for data scientists. In most cases, constructing a complete federated learning job requires only about 6+ lines of Python code. When combined with the Client API (typically 4+ lines), building and running federated learning experiments becomes almost effortless.

Job Recipe

This feature is in technical preview. Not all examples and code have been converted to use Job Recipe yet. However, you can directly experience the recipe with recipe tutorial notebook Job Recipe Tutorials or read the NVFlare Job Recipe, more than half a dozen ready-to-use recipes are provided: Quick Start Series

Enhanced Communication: Port Consolidation and new HTTP Driver

Port Consolidation Previously, FLARE’s server required two separate ports: one for FL client/server communication and another for Admin client/server communication. In 2.7, these are merged into a single configurable port, reducing network configuration complexity. Dual-port mode remains available for environments with stricter network policies.

New HTTPS Driver The HTTP driver has been rewritten using aiohttp to address prior performance limitations. It now matches gRPC performance, while maintaining the same API, TLS support, and backward compatibility with existing deployments.

• Consolidated Port: Reduced from two ports to a single port, simplifying deployment.

• Standard Port Compatibility: Use standard HTTPS port 443 - no need for IT to open additional ports

• High Performance: New HTTP driver matches gRPC in speed and reliability.

Why it matters

Faster Deployment: Eliminates network configuration delays and IT approvals for opening custom ports. FLARE 2.7.0 simplifies network requirements and allows fully secure deployments on standard infrastructure. Check out FL server port consolidation details.

Security Enhancement

Fixed the following issues:

• Unsafe Deserialization - torch.jit.load is replaced with safe-tensor based implementation

• Unsafe Deserialization - Function Call - FOB auto-registration is removed. A whitelist of FOBs is auto-registered.

• Command Injection via Grep Parameters - commands are reimplemented to avoid command injections

Security Enhancements

Many similar issues are also fixed

Develop Edge Applications with FLARE

FLARE 2.7 extends federated learning to edge devices with features that directly address the unique challenges of edge environments:

Scalability: Hierarchical federated architecture Hierarchical FLARE allows millions of edge devices to participate efficiently without connecting each directly to the server.

Intermittent Device Participation: Asynchronous FL based on FedBuff Edge Device Training (Jetson / GPU) handles devices that may join, leave, or fail to return local training results due to network or power interruptions.

Cross-Platform & No Device Programming Required: Data scientists can deploy models to iOS and Android FLARE Mobile Development without writing Swift, Objective-C, Java, or Kotlin. FLARE handles PyTorch → Executorch conversion and device training code automatically.

Simulation Tools: Device simulator for large scale testing (Device Simulation)

FLARE Edge

Try FLARE edge development following the edge examples

Self-Paced-Training Tutorials

Welcome to the five-part course on Federated Learning with NVIDIA FLARE! This course covers everything from the fundamentals to advanced applications, system deployment, privacy, security, and real-world industry use cases.

Federated Learning with NVIDIA FLARE

This tutorial has 100+ notebooks and 80 videos. See details in Self-Paced-Training Tutorials

Extra Features

There are additional new features released in version 2.7.0, including memory management improvements with FileDownloader for large model streaming. You can find more details in Extra Features in v2.7.0.