Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
153 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

End-to-End Evaluation of Federated Learning and Split Learning for Internet of Things (2003.13376v2)

Published 30 Mar 2020 in cs.CR, cs.DC, and cs.LG

Abstract: This work is the first attempt to evaluate and compare felderated learning (FL) and split neural networks (SplitNN) in real-world IoT settings in terms of learning performance and device implementation overhead. We consider a variety of datasets, different model architectures, multiple clients, and various performance metrics. For learning performance, which is specified by the model accuracy and convergence speed metrics, we empirically evaluate both FL and SplitNN under different types of data distributions such as imbalanced and non-independent and identically distributed (non-IID) data. We show that the learning performance of SplitNN is better than FL under an imbalanced data distribution, but worse than FL under an extreme non-IID data distribution. For implementation overhead, we end-to-end mount both FL and SplitNN on Raspberry Pis, and comprehensively evaluate overheads including training time, communication overhead under the real LAN setting, power consumption and memory usage. Our key observations are that under IoT scenario where the communication traffic is the main concern, the FL appears to perform better over SplitNN because FL has the significantly lower communication overhead compared with SplitNN, which empirically corroborate previous statistical analysis. In addition, we reveal several unrecognized limitations about SplitNN, forming the basis for future research.

Citations (174)

Summary

  • The paper demonstrates that SplitNN achieves faster convergence than FL in imbalanced IoT scenarios, while FL excels under extreme non-IID conditions.
  • The paper shows that FL incurs lower communication overhead, making it more suitable for constrained IoT environments despite SplitNN’s memory efficiency on clients.
  • The paper highlights the need for optimizing distributed learning models to support effective and energy-efficient IoT applications on resource-limited devices.

Evaluation of Federated and Split Learning Techniques for IoT Applications

The paper "End-to-End Evaluation of Federated Learning and Split Learning for Internet of Things" thoroughly investigates the applicability of two advanced distributed machine learning techniques—Federated Learning (FL) and Split Neural Network (SplitNN)—in Internet of Things (IoT) contexts. Both techniques offer privacy-preserving benefits as they eliminate the need for raw data transfer to central servers. This paper's novel contribution lies in providing an empirical evaluation comparing these approaches under realistic IoT conditions, where data imbalances and non-IID (non-Independent and Identically Distributed) characteristics are prevalent, and devices are often resource-constrained.

Key Findings

  1. Learning Performance:
    • Federated Learning vs. Split Learning: The paper empirically shows that SplitNN often has superior convergence speed over FL in scenarios with imbalanced data distribution. However, under extreme non-IID data conditions, FL outperforms SplitNN, which occasionally fails to learn.
    • Scalability: FL struggles to maintain model performance as client numbers increase without adjusting for multiple local epochs. SplitNN, however, achieves rapid convergence but struggles with stability and achieving centralized model performance levels.
  2. Implementation Overhead:
    • Extensive testing on Raspberry Pi devices revealed that FL generally resulted in significantly lower communication overhead compared to SplitNN. This factor is crucial for IoT settings where communication resources are limited.
    • Energy and Memory Consumption: SplitNN was advantageous in terms of memory usage due to fewer parameters being trained on the client-side. However, FL proved more efficient in terms of energy consumption due to parallel training across clients.
    • The paper also highlights potential energy savings and reduced training time through ensemble learning strategies for SplitNN, exploiting server-side parallelism capabilities.

Implications and Future Directions

The findings have practical implications suggesting that while FL is preferable under strict communication constraints, SplitNN might be advantageous when computational capabilities of the server can be leveraged, especially with ensemble learning techniques. Moreover, for IoT applications requiring heavy models, neither technique is deemed feasible on resource-constrained devices like Raspberry Pi without substantial optimization.

Further research directions might include developing optimized SplitNN variants for sequential data models like LSTMs and RNNs, which were not addressable by the current paper. Additionally, reducing computation-heavy operations in model architectures could make distributed approaches more viable on IoT devices.

Overall, this comprehensive evaluation provides researchers with a deeper understanding of the trade-offs involved in deploying FL and SplitNN in real-world IoT scenarios and offers a benchmark for subsequent optimizations and applications of these distributed learning paradigms.