DNN_Splitting

Goal-oriented DNN splitting for edge-assisted CNN inference.

This repository contains the reference implementation of the resource allocation strategy presented in:

• F. Binucci, M. Merluzzi, P. Banelli, E. C. Strinati, and P. Di Lorenzo, “Enabling Edge Artificial Intelligence via Goal-oriented Deep Neural Network Splitting,”
  Proc. 19th International Symposium on Wireless Communication Systems (ISWCS), Rio de Janeiro, Brazil, 2024, pp. 1–6.
  DOI: https://doi.org/10.1109/ISWCS61526.2024.10639178
• Preprint: https://arxiv.org/abs/2312.03555

Overview

Deep Neural Network (DNN) splitting is a key enabler of edge AI: it allows the end device to execute an initial portion of the model and offload intermediate features to a nearby Edge Cloud Server (ECS) to complete inference. This creates additional degrees of freedom to balance energy consumption, latency, and task accuracy, while also supporting privacy- and trust-aware deployments.

In this work, we study DNN splitting for edge inference in 6G scenarios from a goal-oriented perspective. In particular, we account for accuracy degradation due to the splitting point (SP) selection and wireless channel conditions, and we propose a dynamic algorithm that jointly controls:

• SP selection,
• local computing resources,
• uplink transmit power, and
• bandwidth allocation,

to achieve a target goal-effectiveness (e.g., minimize energy under latency and accuracy constraints).

Usage

The implementation is organized around two main classes:

• ServerSimulator: simulates ECS-side inference according to the probabilistic model described in the paper.
• DeviceSimulator: simulates the user equipment (UE) resource allocation strategy for edge-assisted DNN splitting.

To run a simulation in a specific scenario, use:

• Simulate.m — update this script to modify the scenario parameters and constraints.

Plotting

To reproduce the figures from the paper:

• Fig. 3: run plotEnergyAccuracyTOForFixedSNRComparison.m and select option 5.
• Fig. 4: run plotEnergyAccuracyTOForFixedSPComparison.m and select option 7.

Citation

If you use this code in your work, please cite:

@inproceedings{binucci2024goalorientedDNNSplitting,
  author    = {Binucci, Francesco and Merluzzi, Mattia and Banelli, Paolo and Strinati, Emilio Calvanese and Di Lorenzo, Paolo},
  title     = {Enabling Edge Artificial Intelligence via Goal-oriented Deep Neural Network Splitting},
  booktitle = {2024 19th International Symposium on Wireless Communication Systems (ISWCS)},
  year      = {2024},
  pages     = {1--6},
  doi       = {10.1109/ISWCS61526.2024.10639178}
}

Contributing

Contributions are welcome.

• Open an issue for bugs/feature requests
• Submit a PR with a clear description and minimal reproducible example
• Add tests for new functionality where possible

License

Licensed under the MIT License. See LICENSE.

Contact

• Maintainer: Francesco Binucci
• Email: francesco.binucci@cnit.it