Abstract
The aggregation of residents' private data drives improvements in smart home services, yet it often necessitates a compromise on privacy. Because users have varying privacy requirements—which are themselves sensitive information—smart homes require a preservation scheme that meets individual demands without a trusted central authority. We propose a novel personalized local differential privacy (PLDP) scheme. In this framework, users preserve data privacy locally by specifying their own confidence levels, reporting only single data elements to collectors. Our scheme retains high data utility while providing rigorous, personalized privacy guarantees.
Introduction
Smart home ecosystems rely on continuous data aggregation—such as tracking temperature preferences or resident demographics—to optimize services and energy efficiency. However, traditional data collection poses significant privacy risks. While Differential Privacy (DP) is considered the gold standard for privacy preservation, its application in smart home environments faces three major hurdles:
- The Trust Gap: Traditional DP often requires a trusted central aggregator. Residents may not want to trust service providers or third parties with their raw data.
- The "One-Size-Fits-All" Problem: Standard Local Differential Privacy (LDP) applies a uniform privacy budget ($\epsilon$) to all users. This often results in insufficient privacy for some or excessive noise that renders the aggregate data useless for the collector.
- Longitudinal Attacks: Repeated data collection allows adversaries to "average out" noise over time, eventually revealing true values.
Proposed Methodology: Personalized Local Differential Privacy (PLDP)
To address these challenges, we introduce a PLDP framework where privacy protection is executed locally on the resident's sensor device. Key components include:
Security Domains
Users define a "security domain" $[s, e]$ within which they wish to protect their data. For instance, a 23-year-old might set a domain of [20, 30]. This approach allows the collector to know the general range while hiding the exact value, significantly reducing the noise required compared to protecting data across the entire possible spectrum (e.g., ages 0–100).
Two-Stage Binary Randomized Response (TBRR)
The core mechanism involves two stages of randomization:
- Stage 1: Data is mapped to a Bloom Filter, and bits are randomized based on the user's personal privacy budget ($p$).
- Stage 2: A second layer of randomization is applied using a collector-defined parameter ($q$). This "double-shuffling" protects against longitudinal tracking by making it harder to reverse-engineer data through repeated observations.
Privacy Budget Estimation
Because a user's chosen privacy level is itself sensitive, it is not shared directly. We developed a high-precision estimation algorithm that allows the collector to derive aggregate privacy budgets from noisy data, improving final result accuracy.
Experimental Evaluation
The framework was evaluated using synthetic datasets mimicking smart home scenarios: norm-age (Normal distribution of ages), exp-age (Exponential decay), and unif-t1/t2 (Uniform temperature distributions).
Key Results
The PLDP scheme was compared against Google's RAPPOR. On the norm-age dataset, the PLDP-2 variant (utilizing security domains) achieved a Mean Relative Error (MRE) of 0.067, a nearly 4x improvement over RAPPOR's 0.251.
As shown in the age distribution analysis, PLDP-2 maintains the "shape" of the data far more effectively than standard LDP methods. Similar results were observed in temperature preference tracking, where PLDP-2 consistently outperformed other methods in distribution matching.
Discussion and Conclusion
Our findings suggest that privacy is not a binary state but a personal choice. By allowing users to specify their own $\epsilon$, participation in data sharing increases, which paradoxically improves aggregate accuracy for the collector. Furthermore, the use of "security domains" proves that hiding data within a relevant neighborhood provides sufficient privacy while preserving utility.
Future work will focus on testing the system against real-world, non-synthetic smart home datasets and optimizing the computational overhead for low-power IoT microcontrollers to ensure seamless integration into existing smart home hardware.
Summary updated