Private Collaborative Edge Inference via Over-the-Air Computation

TL;DR

The paper tackles private collaborative edge inference where independently trained local models are queried in parallel and fused at the wireless edge. It leverages the superposition property of over-the-air computation to aggregate predictions with reduced bandwidth and enhanced privacy, introducing fusion schemes (Belief Averaging, Weighted Belief Averaging, and Majority Voting) and a DP framework achieved via Gaussian noise and random participation. The methodology includes a transmission strategy with a common projection, channel-inversion scaling, and a scaling factor gamma to satisfy a power constraint, along with a detailed privacy analysis and DP guarantees. Empirical results across eight diverse datasets demonstrate that OAC-based methods outperform orthogonal and fully digital baselines under privacy, with strong statistical significance and favorable scalability; the work also provides comprehensive ablations and practical considerations, and releases public source code for reproducibility.

Abstract

We consider collaborative inference at the wireless edge, where each client's model is trained independently on its local dataset. Clients are queried in parallel to make an accurate decision collaboratively. In addition to maximizing the inference accuracy, we also want to ensure the privacy of local models. To this end, we leverage the superposition property of the multiple access channel to implement bandwidth-efficient multi-user inference methods. We propose different methods for ensemble and multi-view classification that exploit over-the-air computation (OAC). We show that these schemes perform better than their orthogonal counterparts with statistically significant differences while using fewer resources and providing privacy guarantees. We also provide experimental results verifying the benefits of the proposed OAC approach to multi-user inference, and perform an ablation study to demonstrate the effectiveness of our design choices. We share the source code of the framework publicly on Github to facilitate further research and reproducibility.

Figures (8)

• Figure 1: Overview of collaborative inference with guarantees.
• Figure 2: Overview of our framework for collaborative private inference at the wireless edge.
• Figure 3: Box plots for the Macro-F1 score for all the datasets in the case of non-private ($\varepsilon=\infty$, left), moderately private, ($\varepsilon=5$, middle), strongly private ($\varepsilon=1$, right) scenarios.
• Figure 4: Comparison of MV-OAC and MV-Orth with different levels as a function of channel SNR (left), participation probability $p$ (middle) and projection dimensions $d$ (right) on the validation split of CIFAR-10 dataset.
• Figure 5: Comparison of BA-OAC and BA-Orth with different levels as a function of channel SNR (left), participation probability $p$ (middle) and projection dimensions $d$ (right) on the validation split of CIFAR-10 dataset.

Theorems & Definitions (15)

• Remark 1
• Remark 2
• Definition 1
• Definition 2
• Remark 3
• Remark 4
• Theorem 1
• proof
• Lemma 1: Theorem 8 in balle2018improving
• Theorem 2