EUVIMEDOpdag EUVIMED
Open access · CC BY-NC-ND 4.0

Privacy-Preserving Feature Extraction with Differentially Private Contrastive Objectives for Time Series

Na Zou, Sophie Martin

American Journal of Machine Learning · 2025

DOI: 10.71465/ajml3458 ↗

At a glance

Journal article in American Journal of Machine Learning, 2025. Written by Na Zou and 1 co-authors. The full text is freely available (open access, CC BY-NC-ND 4.0).

AbstractBibliografiske oplysningerCiteRelated studies

Abstract (original language)

The proliferation of ubiquitous sensing technologies has resulted in an explosion of time series data generation, ranging from personal health metrics collected by wearable devices to complex industrial telemetry in smart infrastructures. While Deep Learning has demonstrated remarkable utility in extracting latent representations from such high-dimensional temporal data, the deployment of these models raises severe privacy concerns. In particular, the inversion of feature embeddings can inadvertently reveal sensitive attributes of the underlying subjects. This paper introduces a novel framework for Privacy-Preserving Feature Extraction using Differentially Private Contrastive Learning (DP-CL). We propose a rigorous methodology that integrates gradient perturbation mechanisms directly into the optimization of a normalized temperature-scaled cross-entropy loss (NT-Xent) adapted for temporal data. By enforcing Differential Privacy (DP) guarantees during the pre-training phase, we ensure that the resulting encoder yields representations that are robust against membership inference and reconstruction attacks while maintaining high downstream utility. Our approach addresses the specific challenges of applying DP to contrastive objectives, such as the sensitivity of gradient magnitudes in projection heads and the impact of noise on alignment and uniformity metrics. Extensive experiments on standard multivariate time series benchmarks demonstrate that our method achieves a superior trade-off between privacy budgets and classification accuracy compared to traditional auto-encoding baselines and supervised differentially private training.

Cite

APA

Zou, N., & Martin, S. (2025). Privacy-Preserving Feature Extraction with Differentially Private Contrastive Objectives for Time Series. American Journal of Machine Learning. https://doi.org/10.71465/ajml3458

Vancouver

Zou N, Martin S. Privacy-Preserving Feature Extraction with Differentially Private Contrastive Objectives for Time Series. American Journal of Machine Learning. 2025. doi:10.71465/ajml3458

BibTeX
@article{Zou2025,
  author = {Zou, Na and Martin, Sophie},
  title = {Privacy-Preserving Feature Extraction with Differentially Private Contrastive Objectives for Time Series},
  journal = {American Journal of Machine Learning},
  year = {2025},
  doi = {10.71465/ajml3458}
}
Start for free →