Abstract
In this paper, we propose Acies, a differential privacy based privacy-preserving classification system for edge computing to secure the classification models offloaded to edge devices. Acies supports popular classifiers such as Nearest Neighborhood, Support Vector Machine and Sparse Representation Classifier with a variety of feature selection methods. According to our evaluation on different datasets, classification models with Acies can be private and remain high utility. Acies achieves reliable privacy protection under reconstruction attacks with minimal impact on classification accuracy (2%-5%) only. Acies outperforms the naive input dataset perturbation methods by up to 30% higher classification accuracy when the privacy requirements of the applications is high ($ϵ$ is less than 2).
Introduction
Edge computing shifts data processing from central cloud servers to local nodes like smartphones and IoT gateways to reduce latency. However, these edge nodes are often located in untrusted environments, making them vulnerable to physical tampering or hacking. If a classification model—such as a face recognizer—is offloaded to an insecure edge device, a malicious actor could perform a reconstruction attack. By analyzing the model's internal logic or the features it extracts, an attacker can reverse-engineer sensitive training data, potentially recreating private images or medical signals.
Traditional solutions often rely on "input perturbation," adding noise to raw data before it reaches the model. While this protects privacy, it frequently degrades model accuracy to the point of uselessness. Acies (Latin for "sharp edge") addresses this by injecting differential privacy noise into the feature extraction process itself, rather than the raw input.
Methodology: Feature Extraction Perturbation
Acies utilizes Differential Privacy (DP) to create a secure "lens" through which the edge device views data. While standard DP approaches add noise to the raw dataset $X$ (input perturbation), Acies perturbs the transformation matrix used for feature extraction.
In methods like SVD, Eigenface, or Fisherface, a transformation matrix $R$ is used to reduce dimensionality and extract key features. Acies adds Laplace noise directly to this matrix $R$. By protecting the transformation process, Acies ensures that the features extracted by the edge device satisfy mathematical privacy requirements while preserving the underlying patterns necessary for high-utility classification.
Experimental Evaluation
The researchers evaluated Acies using two distinct datasets to demonstrate its versatility:
- Extended Yale Face Database B (YaleB): A benchmark for face recognition used to test resistance against the reconstruction of human faces under varying lighting conditions.
- WiFi Channel State Information (CSI): A dataset for human identification based on gait (walking patterns), representing a modern IoT use case where WiFi signals act as sensors.
Key Results and Insights
Acies demonstrates that it is possible to maintain high utility even under strict privacy constraints. When the privacy budget $ϵ$ is low (less than 2), traditional input perturbation methods see accuracy plummet to 30-50%. In contrast, Acies maintains high performance with only a 2-5% drop in accuracy.
Beyond accuracy, Acies provides robust protection against reconstruction attacks. As shown in the evaluation, an attacker's ability to reconstruct original data—measured by Mutual Information (MI)—drops significantly faster with Acies than with traditional methods as noise levels increase. Furthermore, because Acies utilizes dimensionality reduction (compressing features to as little as 5% of their original size), it can speed up classification by up to 13x, making it ideal for low-power IoT devices.
Conclusion
Acies offers a significant advancement for privacy-preserving edge computing by shifting the focus of perturbation from the "object" (data) to the "lens" (feature extraction). While the current framework focuses on linear transformations like PCA and SVD, it successfully supports popular classifiers like kNN and SVM. Future research may extend these principles to non-linear models and deep neural networks to further harden edge-based AI against sophisticated membership inference and reconstruction attacks.
Summary updated