Advances in APPFL: A Comprehensive and Extensible Federated Learning Framework

Zilinghan Li, Shilan He, Ze Yang, Minseok Ryu, Kibaek Kim, Ravi K. Madduri

Abstract

The Advanced Privacy-Preserving Federated Learning (APPFL) framework is an open-source, modular ecosystem designed to bridge the gap between theoretical Federated Learning (FL) research and robust, real-world deployments. APPFL addresses the "full spectrum" of FL challenges, including hardware heterogeneity, data security, and communication efficiency for large-scale models.

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models across distributed data sources without compromising data privacy. However, practical implementation is often hindered by three types of heterogeneity: data (non-IID distributions), computation (varying device speeds), and communication (bandwidth limitations). While many frameworks exist, they often fail to provide a comprehensive solution that integrates advanced scheduling, enterprise-grade security, and high-efficiency communication.

Comparison of APPFL with other popular FL frameworks across various features, highlighting its comprehensive nature.

As illustrated in the comparison above, APPFL distinguishes itself by offering built-in solutions for asynchronous FL, model compression, and institutional-grade authentication—features that are frequently missing or only partially implemented in other popular frameworks like Flower or FedML.

The APPFL Architecture

The APPFL framework introduces a redesigned architecture centered on modularity and extensibility. It consists of three primary layers:

  1. Server and Client Agents: These components act as the system's "brains," managing the core training logic, including local updates and global aggregation strategies.
  2. Communicator Layer: This versatile component supports multiple protocols. Notably, it offers a server-driven communication mode via Globus Compute, allowing servers to dispatch tasks to remote endpoints without requiring clients to open inbound network ports, a common barrier in enterprise environments.
  3. Advanced Schedulers: APPFL implements sophisticated scheduling algorithms like FedCompass. This computing power-aware scheduler dynamically adjusts local training steps based on real-time client performance, achieving high resource utilization (~90%) without the accuracy degradation often seen in purely asynchronous systems.
The modular architecture of APPFL, illustrating the interaction between Server Agents, Client Agents, and the Communicator.

Efficiency and Security

To support large-scale models such as Foundation Models, APPFL integrates ProxyStore, which decouples control signaling from heavy data transfers. This allows the framework to bypass standard protocol message size limits (often as low as 10MB) by using external data connectors like S3 buckets. Additionally, the framework incorporates lossy and lossless compressors (SZ2, SZ3, ZFP), which can reduce model update sizes by 3x to 5x with minimal impact on final accuracy.

Security in APPFL is multi-layered. Beyond algorithmic privacy via Differential Privacy, the framework integrates Globus Auth for institutional-grade identity and access management (IAM). This ensures that only authorized collaborators can participate in a federation, providing a level of security necessary for sensitive sectors like healthcare and finance.

Evaluation and Results

The framework was rigorously evaluated using diverse benchmarks, including CIFAR-10, MNIST, and real-world medical datasets from the FLamby suite (e.g., Fed-TCGA-BRCA, Fed-Heart-Disease). Key findings include:

  • Resource Utilization: The FedCompass scheduler improved hardware utilization to approximately 90% in heterogeneous environments, significantly outperforming traditional synchronous methods that leave faster devices idle.
  • Communication Efficiency: Lossy compression achieved significant data reduction (3-5x) while maintaining model performance, crucial for training over limited bandwidth.
  • Versatility: APPFL demonstrated successful adaptation to Vertical FL (different features, same IDs), Hierarchical FL (using edge servers), and Decentralized FL variants with minimal architectural changes.

Conclusion and Future Work

APPFL provides a robust, extensible foundation for deploying Federated Learning in diverse real-world scenarios. Future development will focus on integrating advanced cryptographic methods such as Secure Multi-Party Computation (SMPC) and Homomorphic Encryption (HE), as well as hardware-level security through Trusted Execution Environments (TEEs). The team also aims to further optimize the framework for fine-tuning massive foundation models across distributed data silos.

Summary updated

Abstract

Federated learning (FL) is a distributed machine learning paradigm enabling collaborative model training while preserving data privacy. In today's landscape, where most data is proprietary, confidential, and distributed, FL has become a promising approach to leverage such data effectively, particularly in sensitive domains such as medicine and the electric grid. Heterogeneity and security are the key challenges in FL, however; most existing FL frameworks either fail to address these challenges adequately or lack the flexibility to incorporate new solutions. To this end, we present the recent advances in developing APPFL, an extensible framework and benchmarking suite for federated learning, which offers comprehensive solutions for heterogeneity and security concerns, as well as user-friendly interfaces for integrating new algorithms or adapting to new applications. We demonstrate the capabilities of APPFL through extensive experiments evaluating various aspects of FL, including communication efficiency, privacy preservation, computational performance, and resource utilization. We further highlight the extensibility of APPFL through case studies in vertical, hierarchical, and decentralized FL. APPFL is open-sourced at https://github.com/APPFL/APPFL.
BibTeX
@misc{li2024advances,
  author = {Zilinghan Li and Shilan He and Ze Yang and Minseok Ryu and Kibaek Kim and Ravi K. Madduri},
  title = {{Advances in APPFL: A Comprehensive and Extensible Federated Learning Framework}},
  year = {2024},
  eprint = {2409.11585},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.11585},
}
Figure