Abstract:Cloud-edge collaborative inference approach splits deep neural networks (DNNs) into two parts that run collaboratively on resource-constrained edge devices and cloud servers, aiming at minimizing inference latency and protecting data privacy. However, even if the raw input data from edge devices is not directly exposed to the cloud, state-of-the-art attacks targeting collaborative inference are still able to reconstruct the raw private data from the intermediate outputs of the exposed local models, introducing serious privacy risks. In this paper, a secure privacy inference framework for cloud-edge collaboration is proposed, termed CIS, which supports adaptively partitioning the network according to the dynamically changing network bandwidth and fully releases the computational power of edge devices. To mitigate the influence introduced by private perturbation, CIS provides a way to achieve differential privacy protection by adding refined noise to the intermediate layer feature maps offloaded to the cloud. Meanwhile, with a given total privacy budget, the budget is reasonably allocated by the size of the feature graph rank generated by different convolution filters, which makes the inference in the cloud robust to the perturbed data, thus effectively trade-off the conflicting problem between privacy and availability. Finally, we construct a real cloud-edge collaborative inference computing scenario to verify the effectiveness of inference latency and model partitioning on resource-constrained edge devices. Furthermore, the state-of-the-art cloud-edge collaborative reconstruction attack is used to evaluate the practical availability of the end-to-end privacy protection mechanism provided by CIS.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is how to protect data privacy while minimizing inference latency during cloud - edge collaborative inference. Specifically, the paper focuses on: 1. **Existing challenges**: Although cloud - edge collaborative inference can reduce inference latency and protect data privacy by splitting the deep neural network (DNN) into two parts to run on edge devices and the cloud respectively, current attack methods can still reconstruct the original private data from the intermediate output results, bringing serious privacy risks. 2. **Proposed solution**: For this reason, the paper proposes a new secure - privacy inference framework CIS (Collaborative Inference Shield), aiming to maximize privacy strength while minimizing the impact on DNN accuracy, and effectively balance the contradiction between availability and privacy in cloud - edge collaborative inference. ### Main contributions - **Adaptive model splitting**: CIS supports adaptively splitting the network according to the dynamically changing network bandwidth to fully utilize the computing power of edge devices and effectively balance the total inference latency and the privacy of edge - sensitive data. - **Differential privacy protection mechanism**: CIS achieves differential privacy protection by adding finely - tuned noise to the intermediate - layer feature maps uploaded to the cloud. At the same time, under the given total privacy budget, it reasonably allocates the rank sizes of feature maps generated by different convolutional filters, making cloud - side inference robust to perturbed data, thus effectively balancing the conflict between privacy and availability. - **Practical evaluation**: A real cloud - edge collaborative inference computing scenario was constructed to verify the effectiveness of inference latency and model splitting on resource - constrained edge devices. In addition, the practical availability of the end - to - end privacy protection mechanism provided by CIS was evaluated using the state - of - the - art cloud - edge collaborative reconstruction attack. ### Formula summary To ensure the effectiveness of differential privacy protection, the following key formulas are introduced in the paper: - **Differential privacy definition**: \[ \Pr[M(D)\in S]\leq e^{\epsilon}\Pr[M(D')\in S] \] where \(D\) and \(D'\) are adjacent data sets, \(M\) is a random mechanism, and \(\epsilon\) is the privacy protection budget. - **Global sensitivity**: \[ \Delta s(f,\|\cdot\|)=\max_{d(D,D') = 1}\|f(D)-f(D')\| \] where \(\|\cdot\|\) represents a distance metric, usually the \(l_1\) or \(l_2\) norm. - **Laplace mechanism**: \[ M_L((x,f(\cdot),\epsilon))=f(x)+(Y_1,Y_2,\cdots,Y_k) \] where \(Y_i\) is a random variable independently and identically distributed from the Laplace distribution \(\text{Lap}\left(\frac{\Delta s}{\epsilon}\right)\). Through these methods, CIS can maintain high inference accuracy and low latency while protecting privacy, thus solving the key challenges in cloud - edge collaborative inference.

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

Privacy-Preserving Collaborative Deep Learning with Unreliable Participants.

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

Data-Driven Caching with Users' Local Differential Privacy in Information-Centric Networks

Extendable Multi-Device Collaborative Pipeline Parallel Inference in the Edge-Cloud Scenario

Edge-based differential privacy computing for sensor–cloud systems

Over-the-Air Collaborative Inference with Feature Differential Privacy

Collaborative Inference over Wireless Channels with Feature Differential Privacy

Learning from Differentially Private Neural Activations with Edge Computing

Joint Content and Communication Resource Allocation for Privacy-Preserving Distributed Collaborative Edge Caching

SecoInfer: Secure DNN End-Edge Collaborative Inference Framework Optimizing Privacy and Latency

Differentially Private Federated Tensor Completion for Cloud-Edge Collaborative AIoT Data Prediction

EdgeSanitizer: Locally Differentially Private Deep Inference at the Edge for Mobile Data Analytics

PCNNCEC: Efficient and Privacy-Preserving Convolutional Neural Network Inference Based on Cloud-Edge-Client Collaboration

Privacy-Preserving Resource Management for Distributed Collaborative Edge Caching Systems

Pratical Privacy-preserving Convolutional Neural Network Inference Framework with Edge Computing for Health Monitoring

Model Protection: Real-Time Privacy-Preserving Inference Service for Model Privacy at the Edge

A privacy protection approach in edge-computing based on maximized dnn partition strategy with energy saving

PATROL: Privacy-Oriented Pruning for Collaborative Inference Against Model Inversion Attacks

A survey on privacy inference attacks and defenses in cloud-based Deep Neural Network

Privacy preserving data sharing and analysis for edge-based architectures