Abstract
Federated Learning (FL) has emerged as a critical paradigm for training machine learning models while maintaining data privacy. However, its application is often limited to high-performance edge devices like smartphones. This paper proposes a scalable, multi-tier serverless architecture designed to extend FL capabilities to low-power IoT and edge sensors. By utilizing mobile phones as intermediate compute relays and serverless functions for cloud orchestration, the framework overcomes hardware constraints and connectivity issues inherent in tiny sensors while strictly preserving data privacy at the edge.
Introduction
The proliferation of Internet of Things (IoT) devices—from wearables to smart home sensors—generates vast amounts of sensitive data. Traditional machine learning approaches often require centralizing this data in the cloud, raising significant privacy concerns and consuming substantial bandwidth. Federated Learning offers a solution by training models locally on user devices. While effective for smartphones, FL faces significant hurdles when applied to the broader IoT ecosystem.
The Problem: IoT Resource and Connectivity Constraints
Most existing FL research assumes devices have sufficient CPU, memory, and stable internet connectivity. Low-power IoT devices, however, suffer from two major hurdles:
- Resource Constraints: They lack the computational power and memory required to execute complex machine learning training algorithms.
- Connectivity Issues: Many sensors lack direct high-bandwidth connections to the cloud, often relying on low-energy local protocols like BLE.
Current workarounds often involve transmitting raw data to the cloud, which defeats the privacy-preserving purpose of FL. This research addresses the need for a scalable infrastructure that allows resource-constrained IoT devices to participate in the FL process without compromising data security.
Proposed Architecture
The authors propose a multi-tier architecture that bridges the gap between simple sensors and sophisticated cloud services. The system is designed to be elastic, cost-effective, and easily deployable on public cloud infrastructure.
The Mobile Relay Layer
A central innovation of this architecture is the use of the smartphone as a "Federation Node." Instead of communicating directly with the cloud, IoT hardware boards connect to a nearby Android phone via low-energy protocols such as BLE or REST. The smartphone acts as a compute-heavy proxy, performing the actual model training on the data collected from the sensors. This "last mile" solution ensures that raw sensor data never leaves the local environment.
Serverless Cloud Orchestration
In the cloud, the system leverages event-driven serverless services (e.g., AWS Lambda or Azure Functions) to manage the FL lifecycle. This serverless approach offers horizontal scalability and cost efficiency, as resources are only consumed when devices are active. Key orchestration tasks include:
- Device Monitoring & Selection: Tracking available mobile relays and selecting a diverse subset for each training round.
- Data Aggregation: Implementing the "Federated Averaging" (FedAvg) algorithm to combine model updates from various relays.
- Model Management: Pushing updated global models back to the edge devices.
Machine Learning and Storage Pipeline
The heavy lifting of global model aggregation and training is handled by a dedicated service running TensorFlow Federated (TFF) on an independent Virtual Machine (IaaS). This is supported by a geo-replicated database for maintaining device states and "Hot" Blob Storage for high-speed model data exchange.
Discussion and Key Insights
The proposed architecture offers several advantages:
- Democratization of FL: By using serverless components, the infrastructure becomes more accessible to developers, reducing the barrier to entry for deploying FL systems.
- Elastic Scalability: The serverless nature allows the system to scale horizontally to accommodate a growing number of IoT devices without manual intervention.
- Privacy at the Edge: The multi-tier approach ensures that sensitive data, such as heart rate or audio feeds, remains local to the user's personal devices.
Limitations and Future Work
As an architectural design proposal, this work currently lacks empirical validation. Future research must address:
- Performance Metrics: Quantifying the latency of the mobile-to-cloud relay and the impact on smartphone battery life.
- Data Heterogeneity: Developing strategies to handle "Non-IID" data distributions across different users, a classic challenge in FL.
- Proximity Dependency: Addressing scenarios where IoT devices are isolated from capable mobile relays.
Conclusion
This paper presents a viable architectural blueprint for integrating low-energy IoT devices into the Federated Learning ecosystem. By combining mobile relays with serverless cloud orchestration, the framework enhances machine learning capabilities at the edge while maintaining a strict posture on data privacy.
Summary updated