[Paper Review] Lightweight and Unobtrusive Data Obfuscation at IoT Edge for Remote Inference

Dixing Xu, Mengyao Zheng+4 more|arXiv (Cornell University)|Dec 20, 2019

Privacy-Preserving Technologies in Data32 references6 citations

TL;DR

This paper proposes a lightweight, unobtrusive neural network-based data obfuscation technique (ObfNet) at IoT edge devices to protect raw inference data privacy in remote deep learning inference. By applying a small-scale, non-invertible neural transform before transmission, the method preserves backend inference accuracy while effectively concealing original data, with minimal computational and energy overhead on edge devices and negligible latency impact on high-bandwidth networks.

ABSTRACT

Executing deep neural networks for inference on the server-class or cloud backend based on data generated at the edge of Internet of Things is desirable due primarily to the limited compute power of edge devices and the need to protect the confidentiality of the inference neural networks. However, such a remote inference scheme incurs concerns regarding the privacy of the inference data transmitted by the edge devices to the curious backend. This paper presents a lightweight and unobtrusive approach to obfuscate the inference data at the edge devices. It is lightweight in that the edge device only needs to execute a small-scale neural network; it is unobtrusive in that the edge device does not need to indicate whether obfuscation is applied. Extensive evaluation by three case studies of free spoken digit recognition, handwritten digit recognition, and American sign language recognition shows that our approach effectively protects the confidentiality of the raw forms of the inference data while effectively preserving the backend's inference accuracy.

Motivation & Objective

• To address privacy leakage in remote inference systems where edge devices transmit raw data to cloud backends.
• To protect the confidentiality of raw inference data without compromising backend model accuracy.
• To design a solution that imposes minimal computational and energy overhead on resource-constrained IoT edge devices.
• To ensure the obfuscation mechanism is unobtrusive, requiring no explicit signaling or modification of existing inference pipelines.
• To evaluate the approach across diverse real-world applications including speech, handwriting, and sign language recognition.

Proposed method

• A small-scale neural network, ObfNet, is deployed at the IoT edge to transform raw inference data into obfuscated form before transmission.
• ObfNet is trained to preserve the class prediction of the original data under the backend inference model (InfNet), ensuring I(O(x)) ≈ I(x) for most inputs.
• The obfuscation is non-invertible and non-linear, making it difficult for a curious backend to recover the original data.
• The method operates without signaling or modification of the backend inference pipeline, ensuring unobtrusiveness.
• The approach is evaluated using three real-world case studies: free spoken digit recognition (FSD), MNIST handwritten digit recognition, and American Sign Language (ASL) recognition.
• Performance is measured in terms of inference accuracy preservation, energy consumption, execution time, and data transmission overhead.

Experimental results

Research questions

• RQ1 Can a lightweight neural obfuscation network at the edge preserve the accuracy of remote deep learning inference while protecting raw data privacy?
• RQ2 How does the proposed obfuscation method perform in terms of computational and energy overhead on edge devices?
• RQ3 To what extent does the obfuscation prevent reconstruction or inference of original data by a curious backend?
• RQ4 How does the remote inference latency and energy efficiency compare to local inference on edge devices?
• RQ5 Can the method be applied effectively across diverse sensing modalities such as audio, image, and gesture recognition?

Key findings

• The proposed ObfNet approach preserves backend inference accuracy across all three case studies—FSD, MNIST, and ASL—despite obfuscation of raw data.
• On Coral TPU, ObfNet incurred only 20 mJ, 2 mJ, and 9.5 mJ of energy per sample for FSD, MNIST, and ASL, respectively, demonstrating minimal energy overhead.
• Per-sample inference time on Jetson was 14.318 ms (FSD), 8.351 ms (MNIST), and 100.467 ms (ASL), showing low latency on high-performance edge hardware.
• Data transmission times were negligible, with less than 1 ms at 15 Mbps, making remote inference faster than local inference when network speed exceeds 15 Mbps.
• Remote inference achieved lower total time delay and better power efficiency than local inference on edge devices when using high-bandwidth connections (e.g., 4G > 100 Mbps).
• Volunteers could not distinguish obfuscated audio samples by ear, and visual inspection showed no meaningful recovery of original images, confirming strong privacy protection.