Abstract:Collaborative inference among multiple wireless edge devices has the potential to significantly enhance Artificial Intelligence (AI) applications, particularly for sensing and computer vision. This approach typically involves a three-stage process: a) data acquisition through sensing, b) feature extraction, and c) feature encoding for transmission. However, transmitting the extracted features poses a significant privacy risk, as sensitive personal data can be exposed during the process. To address this challenge, we propose a novel privacy-preserving collaborative inference mechanism, wherein each edge device in the network secures the privacy of extracted features before transmitting them to a central server for inference. Our approach is designed to achieve two primary objectives: 1) reducing communication overhead and 2) ensuring strict privacy guarantees during feature transmission, while maintaining effective inference performance. Additionally, we introduce an over-the-air pooling scheme specifically designed for classification tasks, which provides formal guarantees on the privacy of transmitted features and establishes a lower bound on classification accuracy.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the privacy risks and communication overhead problems faced by wireless edge devices when transmitting extracted features during collaborative inference. Specifically: 1. **Privacy protection**: In traditional wireless collaborative inference methods, the extracted features may expose sensitive personal information during transmission, leading to privacy leakage. To solve this problem, the paper proposes a new privacy - protection mechanism - Feature Differential Privacy (FDP), which ensures strict protection of feature privacy during transmission. 2. **Communication efficiency**: In multi - device collaborative inference, a large amount of feature data transmission will lead to high communication overhead. For this reason, the paper designs an optimized transmission scheme to improve communication efficiency by reducing the amount of communication while maintaining effective inference performance. 3. **Classification accuracy**: To ensure the accuracy of classification tasks, the paper introduces an over - the - air pooling scheme, which not only provides strict privacy protection but also establishes a lower bound for classification accuracy. ### Main contributions of the paper 1. **Introduction of the Feature Differential Privacy framework**: A new privacy framework, called Feature Differential Privacy (FDP), is proposed to protect the privacy of target features during feature transmission. 2. **Theoretical analysis and experimental verification**: The theoretical lower bound of classification accuracy is provided, and the effectiveness of the proposed framework is verified through numerical experiments. 3. **Two private feature - aware transmission schemes**: Two transmission schemes considering feature quality are proposed. They improve classification accuracy through entropy measurement but will increase a certain privacy leakage risk. 4. **First study on the privacy problem of collaborative inference on wireless channels**: It fills the research gap in the existing literature on the privacy - utility trade - off in the inference stage and provides theoretical insights and practical verification. ### Summary By introducing Feature Differential Privacy and an optimized transmission scheme, this paper solves the privacy risks and communication overhead problems in wireless collaborative inference, while ensuring the efficiency and accuracy of classification tasks.

Collaborative Inference over Wireless Channels with Feature Differential Privacy

Over-the-Air Collaborative Inference with Feature Differential Privacy

Privacy-Preserving Collaborative Deep Learning with Unreliable Participants.

A Distributed Privacy-Preserving Framework for Deep Learning with Edge-Cloud Computing.

Private Collaborative Edge Inference via Over-the-Air Computation

NetDP: In-Network Differential Privacy for Large-Scale Data Processing

Privacy-preserving Security Inference Towards Cloud-Edge Collaborative Using Differential Privacy

Covert and Reliable Semantic Communication Against Cross-Layer Privacy Inference over Wireless Edge Networks

Performance-Enhanced Federated Learning with Differential Privacy for Internet of Things

PrivStream: A Privacy-Preserving Inference Framework on IoT Streaming Data at the Edge

A Two-Stage Differential Privacy Scheme for Federated Learning Based on Edge Intelligence

Guaranteeing Differential Privacy in Distributed Fusion Estimation

Differentially Private Over-the-Air Federated Learning Over MIMO Fading Channels

QI-DPFL: Quality-Aware and Incentive-Boosted Federated Learning with Differential Privacy

Hybrid Differential Privacy Based Federated Learning for Internet of Things

Edge-based differential privacy computing for sensor–cloud systems

PrivStream: Enabling Privacy-Preserving Inferences on IoT Data Stream at the Edge

Privacy Preserving Semi-Decentralized Mean Estimation over Intermittently-Connected Networks

Multilayer Nonlinear Processing for Information Privacy in Sensor Networks

Privacy-preserving multi-channel communication in Edge-of-Things.

FORESEEN: Towards Differentially Private Deep Inference for Intelligent Internet of Things