Privacy-Aware Multi-Device Cooperative Edge Inference with Distributed Resource Bidding

TL;DR

The paper tackles privacy leakage in multi-device cooperative edge inference by formulating the resource bidding and feature compression problem as a DEC-POMDP and solving it with a MADDPG-based algorithm. Each MD independently learns a policy to bid for MEC service and select intermediate feature compression ratios, balancing classification accuracy against privacy leakage quantified by a distortion metric such as SSIM. The approach yields competitive performance against centralized upper bounds and outperforms channel-agnostic baselines, with additional gains when accounting for inference data difficulty. This work enables privacy-preserving, scalable edge inference in heterogeneous, dynamic wireless environments, with implications for practical MEC deployments.

Abstract

Mobile edge computing (MEC) has empowered mobile devices (MDs) in supporting artificial intelligence (AI) applications through collaborative efforts with proximal MEC servers. Unfortunately, despite the great promise of device-edge cooperative AI inference, data privacy becomes an increasing concern. In this paper, we develop a privacy-aware multi-device cooperative edge inference system for classification tasks, which integrates a distributed bidding mechanism for the MEC server's computational resources. Intermediate feature compression is adopted as a principled approach to minimize data privacy leakage. To determine the bidding values and feature compression ratios in a distributed fashion, we formulate a decentralized partially observable Markov decision process (DEC-POMDP) model, for which, a multi-agent deep deterministic policy gradient (MADDPG)-based algorithm is developed. Simulation results demonstrate the effectiveness of the proposed algorithm in privacy-preserving cooperative edge inference. Specifically, given a sufficient level of data privacy protection, the proposed algorithm achieves 0.31-0.95% improvements in classification accuracy compared to the approach being agnostic to the wireless channel conditions. The performance is further enhanced by 1.54-1.67% by considering the difficulties of inference data.

Figures (4)

• Figure 1: Privacy-aware multi-device cooperative edge inference.
• Figure 2: Classification accuracy and SSIM value versus the compression ratio of intermediate features on the CIFAR-10 and STL-10 datasets. Details of the classification models and data reconstruction models are illustrated in Sec. V.
• Figure 3: (a) Averaged cumulative reward achieved by the proposed MADDPG-based algorithm over $5,000$ episodes during the learning process. The weight coefficients of \ref{['eq:reward']} are set as $t_{1, 1}=t_{1, 2}=0.2, t_2=0.8$. (b) Classification accuracy versus the value of $m$.
• Figure 4: Classification accuracy versus SSIM index value with testing datasets. The classification accuracy of local inference and edge inference without feature compression are 73.0% and 90.0% for MD 1, and 72.0% and 88.4% for MD 2, respectively.