Adaptive Lightweight Federated Learning for Resource-Constrained IoT Networks

Project Abstract

In resource-constrained IoT networks, deploying machine learning workflows is hindered by limited computational power, memory, energy, and intermittent connectivity, which together impede robust model performance and timely inference. This work presents an adaptive lightweight federated learning framework designed to address these constraints while preserving data privacy and reducing communication overhead. The approach integrates three core contributions (1) a dynamic model compression and quantization pipeline that adjusts to device capabilities and current energy budgets, enabling on-device local training with minimal resource consumption; (2) an asynchronous, straggler-tolerant federation strategy that leverages partial updates and selective client participation to maintain convergence speed without overburdening any single device or network; and (3) a communication-aware aggregation mechanism augmented with error-resilient encoding and scheduled synchronization that reduces uplink traffic while maintaining global model accuracy. The framework employs a lightweight neural architecture optimized for IoT hardware, paired with a meta-learning-driven scheduler that estimates device readiness and data heterogeneity to determine optimal rounds, client cohorts, and compression settings. A novel convergence analysis under non-iid data and intermittent connectivity demonstrates that the proposed method achieves near-centralized performance under realistic constraints, with theoretical guarantees on regret and communication savings. Empirical evaluation on benchmark IoT datasets and a simulated edge-centric testbed reveals substantial improvements in energy efficiency, communication overhead, and latency, without sacrificing accuracy. Specifically, the framework achieves up to 75% reduction in total communication volume and 40% improvement in wall-clock training time compared to standard federated averaging, while maintaining accuracy within 2–4 percentage points of a full-precision, centralized baseline under diverse non-iid settings. The adaptive compressor dynamically selects quantization levels and model sparsity patterns in response to real-time device energy states and network conditions, effectively balancing local computation and global model quality. The asynchronous federation component tolerates stragglers by prioritizing updates from high-utility clients and leveraging stale-but-informative gradients to accelerate convergence, which is particularly beneficial in heterogeneous IoT ecosystems. The evaluation also investigates robustness to unreliable links, quantization noise, and concept drift, demonstrating graceful degradation and resilience through adaptive replanning of participating devices and trigger-based re-aggregation. By integrating device-aware scheduling, scalable communication, and privacy-preserving aggregation, the proposed framework offers a practical solution for deploying intelligent services in large-scale IoT deployments, such as smart homes, industrial automation, and environmental monitoring. The work contributes a holistic design that unifies model compression, asynchronous federation, and communication-aware aggregation, accompanied by an open-source prototype and a comprehensive benchmark suite to enable reproducibility and further research in resource-constrained federated learning for IoT.

Project Overview

What This Project Is About

---

A straightforward exploration of how to train machine learning models across many small devices (like sensors and gadgets) without sending their data to a central server. The project focuses on making this collaboration lightweight, so devices with limited power, memory, or bandwidth can participate.

---

---

The Problem It Addresses

---

Many IoT systems collect data locally but rely on centralized training, which can raise privacy concerns, drain battery life, and require expensive communication. This project seeks a way for devices to learn from their own data while sharing only small updates, reducing resource use and preserving privacy.

---

---

Objectives of the Project

---

1. Explain the basics of federated learning and why it matters for IoT.
2. Identify the key resource limits in IoT devices and networks.
3. Develop a lightweight training approach that adapts to device constraints.
4. Evaluate performance in terms of accuracy, speed, and energy use.
5. Provide practical guidelines for deploying the method in real IoT settings.

---

---

What You Will Do Step by Step

---

1) Review simple explanations of federated learning and IoT constraints.

2) Design a lightweight learning method that reduces computation and communication.

3) Create a small test setup with simulated or real IoT devices.

4) Train models locally and share tiny updates, then aggregate them.

5) Measure accuracy, energy use, and time, and compare with conventional methods.

6) Adjust the approach to balance speed, accuracy, and resource use.

---

---

Expected Outcome

A practical, easy-to-implement framework that lets resource-limited IoT devices collaboratively learn while using less power and bandwidth, with clear guidance for real-world adoption.