Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Resource Allocation and Secure Wireless Communication in the Large Model-based Mobile Edge Computing System

Zefan Wang, Yitong Wang, Jun Zhao

TL;DR

This work tackles privacy-preserving offsite-tuning of large models in a mobile-edge computing (MEC) setting, where data owners receive a lightweight adapter $\mathcal{A}_n$ and a compressed emulator $\mathcal{E}^*_n$ and return a fine-tuned adapter through a confidential uplink. It formulates a joint resource-allocation problem to maximize the utility-consumption ratio $\mathcal{U}/\mathcal{S}$ under physical-layer security, solved via a hybrid algorithm that combines the Dinkelbach transform, fractional programming, successive convex approximation, and alternating optimization. The solution decomposes into Subproblem 1 (optimizing adapter-extraction rate $\boldsymbol{\phi}$ and the auxiliary parameter $y$) and Subproblem 2 (optimizing bandwidth, power, GPU frequencies, and the time variable), with closed-form and convex reformulations derived through KKT and FP/SCA techniques. Experimental results show the proposed method outperforms baselines in energy, delay, and the overall utility-cost ratio, demonstrating practical viability for privacy-preserving LM fine-tuning at the network edge.

Abstract

With the rapid advancement of large models and mobile edge computing, transfer learning, particularly through fine-tuning, has become crucial for adapting models to downstream tasks. Traditionally, this requires users to share their data with model owners for fine-tuning, which is not only costly but also raises significant privacy concerns. Furthermore, fine-tuning large-scale models is computationally intensive and often impractical for many users. To tackle these challenges, we introduce a system that combines offsite-tuning with physical-layer security, which provides local data owners with a lightweight adapter and a compressed emulator. Data owners then fine-tune the adapter locally and securely send it back to the model owners through a confidential channel for integration, ensuring privacy and resource conservation. Our paper focuses on optimizing computational resource allocation among data owners and the large model owner deployed on edge, and on the compression ratio of adapters. We incorporate a secrecy uplink channel to maximize the utility that we defined while minimizing system costs like energy consumption and delay. The optimization uses the Dinkelbach algorithm, fractional programming, successive convex approximation and alternating optimization. Experiments demonstrate our algorithm's superiority over existing methods.

Resource Allocation and Secure Wireless Communication in the Large Model-based Mobile Edge Computing System

TL;DR

This work tackles privacy-preserving offsite-tuning of large models in a mobile-edge computing (MEC) setting, where data owners receive a lightweight adapter and a compressed emulator and return a fine-tuned adapter through a confidential uplink. It formulates a joint resource-allocation problem to maximize the utility-consumption ratio under physical-layer security, solved via a hybrid algorithm that combines the Dinkelbach transform, fractional programming, successive convex approximation, and alternating optimization. The solution decomposes into Subproblem 1 (optimizing adapter-extraction rate and the auxiliary parameter ) and Subproblem 2 (optimizing bandwidth, power, GPU frequencies, and the time variable), with closed-form and convex reformulations derived through KKT and FP/SCA techniques. Experimental results show the proposed method outperforms baselines in energy, delay, and the overall utility-cost ratio, demonstrating practical viability for privacy-preserving LM fine-tuning at the network edge.

Abstract

With the rapid advancement of large models and mobile edge computing, transfer learning, particularly through fine-tuning, has become crucial for adapting models to downstream tasks. Traditionally, this requires users to share their data with model owners for fine-tuning, which is not only costly but also raises significant privacy concerns. Furthermore, fine-tuning large-scale models is computationally intensive and often impractical for many users. To tackle these challenges, we introduce a system that combines offsite-tuning with physical-layer security, which provides local data owners with a lightweight adapter and a compressed emulator. Data owners then fine-tune the adapter locally and securely send it back to the model owners through a confidential channel for integration, ensuring privacy and resource conservation. Our paper focuses on optimizing computational resource allocation among data owners and the large model owner deployed on edge, and on the compression ratio of adapters. We incorporate a secrecy uplink channel to maximize the utility that we defined while minimizing system costs like energy consumption and delay. The optimization uses the Dinkelbach algorithm, fractional programming, successive convex approximation and alternating optimization. Experiments demonstrate our algorithm's superiority over existing methods.
Paper Structure (19 sections, 1 theorem, 44 equations, 4 figures, 1 algorithm)

This paper contains 19 sections, 1 theorem, 44 equations, 4 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Resource allocation over wireless communication
  4. Large model with mobile edge computing.
  5. Secure wireless communication under physical layer security
  6. System Model and Problem Formulation
  7. System Model
  8. Solution of the joint optimization problem
  9. Subproblem 1
  10. Subproblem 2
  11. Solution to Subproblem 1
  12. Solution to Subproblem 2
  13. Experimental results
  14. Parameter Setting
  15. Performance Comparison with baselines
  16. ...and 4 more sections

Key Result

Theorem 1

The optimal solution of problem $\mathbb{P}_3$ is where $\eta_n$ needs to satisfy with $\widetilde{\phi}_n(\eta_n) |_{a}^{b} := \text{max}(a, \text{min}( \widetilde{\phi}_n(\eta_n) ,b))$.

Figures (4)

  • Figure 1: Optimizing the UCR of a large model system with $N$ data owners and a model owner through joint resource allocation optimization.
  • Figure 2: Metrics concerning the number of users.
  • Figure 3: The system utility-cost ratio (UCR) versus various parameters.
  • Figure 4: Experiments with different weight parameters.

Theorems & Definitions (1)

  • Theorem 1