Learning, Computing, and Trustworthiness in Intelligent IoT Environments: Performance-Energy Tradeoffs

Beatriz Soret, Lam D. Nguyen, Jan Seeger, Arne Bröring, Chaouki Ben Issaid, Sumudu Samarakoon, Anis El Gabli, Vivek Kulkarni, Mehdi Bennis, Petar PopovskiarXiv 2021 · DOI · arXiv

TL;DR

This paper introduces the concept of Intelligent IoT Environments (iIoTe), where heterogeneous devices collaboratively execute autonomous tasks. By analyzing the interplay between distributed learning, edge computing, and blockchain, the authors demonstrate how to navigate the complex tradeoffs between latency, security, and energy consumption in resource-constrained wireless networks.

What Problem Does This Paper Address?

The Internet of Things (IoT) is shifting from a cloud-centric model to an edge-centric one to meet the demands of low-latency, real-time applications like autonomous manufacturing and smart farming. However, this shift introduces three major hurdles:

  1. Complexity: Applications require collaborative sensing, reasoning, and control across diverse devices.
  2. Trust: Devices often belong to different parties, necessitating built-in security without a central authority.
  3. Resource Constraints: Most edge devices are battery-powered and rely on limited wireless bandwidth (like NB-IoT), making energy efficiency a critical bottleneck.

While these areas—Distributed Learning (ML/AI), Edge Computing, and Distributed Ledger Technology (DLT)—are often studied in isolation, this paper addresses the need for a holistic understanding of how they interact and impact the overall energy budget of an iIoTe.

Methods

The authors propose several novel approaches to optimize the three pillars of iIoTe:

  • Distributed Learning: To avoid the energy drain of communicating with a central server, they propose Dynamic GADMM (Group Alternating Direction Method of Multipliers). This leverages spatial sparsity by allowing nodes to only communicate with immediate neighbors. They further refine this into CQ-GGADMM (Censored Quantized Generalized GADMM), which adds temporal sparsity (only sending updates when the model changes significantly) and quantization (reducing the bit-size of updates) to slash communication energy.
  • Edge Computing (Task Allocation): Distributing application components across the network is an NP-hard problem. The authors use Integer Linear Programming (ILP) to model optimal energy usage. Because optimal solvers are too slow for real-time use, they introduce a Linear Heuristic that approximates network costs to find near-optimal solutions rapidly.
  • Trustworthiness (DLT): The paper models a lightweight DLT architecture where IoT devices act as "light nodes," receiving only block headers to minimize storage and energy. They provide a mathematical framework to calculate the end-to-end latency and energy consumption of the consensus process (Proof-of-Work) over an NB-IoT wireless link.
Fig. 1: iIoTe in a manufacturing plant.

Datasets

  • Distributed Learning: The Body Fat dataset from the UCI machine learning repository was used to evaluate the convergence and energy efficiency of the GADMM variants on a linear regression task.
  • Edge Computing: The authors generated synthetic network and application graphs based on the specifications of the Siemens range of industrial computers to test their allocation algorithms.
  • DLT Modeling: Evaluation was based on standard NB-IoT (3GPP) communication parameters and queueing models.

Key Insights and Ideas

  • Sparsity is the Secret to Efficiency: In distributed learning, you don't need every node to talk all the time. By enforcing neighbor-only communication and "censoring" minor updates, the system achieves convergence with a fraction of the energy required by standard Federated Learning.
  • The Heuristic Speedup: In task allocation, the gap between "perfect" and "good enough" is massive in terms of computation time. A heuristic can find a solution in seconds that would take a standard solver days, while still maintaining 60-80% of the optimal energy efficiency.
  • Trust has a "Tax": Implementing DLT provides immutability and security but introduces a clear tradeoff. The consensus process adds latency and communication overhead that must be carefully balanced against the required level of trustworthiness.
  • Holistic Integration: As shown in Figure 9, the iIoTe workflow integrates these technologies: data is processed at the edge, models are trained locally via FL, and results are verified and recorded via DLT, all while an orchestrator optimizes task placement for energy.

Key Results

  • Learning Efficiency: CQ-GGADMM exhibited the lowest total communication energy compared to standard ADMM and GADMM, proving that quantization and censoring are highly effective for wireless IoT.
Loss vs. Sum Energy for different GADMM variants
  • Allocation Performance: For large-scale problems (e.g., 20 nodes), the optimal ILP solver took approximately 10 days (864,104 seconds) to find a solution, whereas the proposed linear heuristic took only 8 seconds.
  • DLT Tradeoffs: The study of DLT over NB-IoT showed that while increasing communication resources initially reduces latency, there is a "tipping point" where the overhead of resource reservation begins to decrease battery life significantly.

Limitations and Future Work

The authors acknowledge that their current models for resource allocation and DLT are primarily static. Future work involves extending these solutions to dynamic environments where devices (like drones or AGVs) move and edge nodes may become intermittently unavailable. Additionally, they aim to investigate the joint optimization of computing and communication resources, rather than treating them as sequential optimization steps, to further push the boundaries of energy efficiency.

Summary updated

Abstract

An Intelligent IoT Environment (iIoTe) is comprised of heterogeneous devices that can collaboratively execute semi-autonomous IoT applications, examples of which include highly automated manufacturing cells or autonomously interacting harvesting machines. Energy efficiency is key in such edge environments, since they are often based on an infrastructure that consists of wireless and battery-run devices, e.g., e-tractors, drones, Automated Guided Vehicle (AGV)s and robots. The total energy consumption draws contributions from multipleiIoTe technologies that enable edge computing and communication, distributed learning, as well as distributed ledgers and smart contracts. This paper provides a state-of-the-art overview of these technologies and illustrates their functionality and performance, with special attention to the tradeoff among resources, latency, privacy and energy consumption. Finally, the paper provides a vision for integrating these enabling technologies in energy-efficient iIoTe and a roadmap to address the open research challenges
BibTeX
@misc{soret2021learning,
  author = {Beatriz Soret and Lam D. Nguyen and Jan Seeger and Arne Bröring and Chaouki Ben Issaid and Sumudu Samarakoon and Anis El Gabli and Vivek Kulkarni and Mehdi Bennis and Petar Popovski},
  title = {{Learning, Computing, and Trustworthiness in Intelligent IoT Environments: Performance-Energy Tradeoffs}},
  year = {2021},
  eprint = {2110.01686},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.01686},
}
Figure