SCEM: A Structure-Preserving Privacy Framework for Sensor-Generated Time-Series Data

Area of Science:

• Data Security
• Machine Learning
• Sensor Networks

Background:

• Intelligent sensor systems and IoT platforms generate valuable time-series data.
• Data characteristics can inadvertently reveal sensitive operational details.
• Existing methods may not adequately protect feature-level information.

Purpose of the Study:

• To introduce Structured Constrained Energy Mapping (SCEM), a general framework for feature-selective protection of sensor time-series data.
• To address the challenge of sensitive information leakage from time-series data.
• To provide a customizable solution for protecting heterogeneous sensitive functionals.

Main Methods:

• SCEM employs a four-stage workflow: feature mapping, structured perturbation, feature recovery, and utility-oriented calibration.
• A generalized privacy energy representation facilitates the transition from mapping to perturbation.
• The framework supports customizable protection mechanisms through a unified instantiation interface.

Main Results:

• SCEM achieved a low unweighted mean attack success rate (ASR@10%) of 4.69% across 18 feature-dataset pairs.
• Experiments demonstrated reduced recoverability of predefined feature-level information under direct feature inference attacks.
• SCEM maintained competitive forecasting utility compared to baselines and original data.

Conclusions:

• SCEM effectively reduces feature-level leakage in sensor time-series data.
• The framework offers a favorable privacy-utility trade-off, preserving temporal structures for analysis.
• SCEM serves as a practical, model-free preprocessing layer for secure data sharing.