Balancing Utility, Privacy, and Energy in Internet of Things SystemsPötter, Henrique (2023) Balancing Utility, Privacy, and Energy in Internet of Things Systems. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractAs the Internet of Things (IoT) enters consumer markets, smart devices with diverse sensing capabilities and always-on connectivity have become more accessible to the public. These devices bring automation and data-driven insights, but their widespread presence increases the risk of exposing private or confidential information. A common approach to protect privacy is using privacy-preserving solutions such as data obfuscation, but this approach has drawbacks. It might bolster privacy but also compromise data's utility. Also, it can demand additional energy, affecting the mobility of battery or energy-harvesting IoT device deployments. This dissertation contributes to designing, implementing, and evaluating energy-efficient utility-aware privacy solutions to enable IoT systems to protect privacy and improve reliability. We study the IoT Utility, Privacy, and Energy (UPE) tradeoffs in three phases: to (1) define the requirements for privacy solutions to better balance the UPE tradeoffs; (2) understand the limitations of privacy solutions in the context of federated learning applications; and (3) preserve user privacy through the selective removal of only the sensitive contents of data. In the first phase, we develop a new methodology to evaluate the UPE tradeoffs of privacy-preserving techniques by augmenting the conventional Utility-Privacy problem by adding energy consumption. This model is evaluated with two data modalities: image classification and audio applications. In phase two, we develop a methodology to assess the privacy guarantees of neural network inferences using differential privacy with federated learning for IoT. Lastly, in the third phase, we seek to minimize energy consumption by developing a solution to only target the most sensitive data segments. Here we create PrivSpeech, a framework that uses a lightweight neural network that only obfuscates the sensitive attributes while maintaining the utility with minimal energy consumption. We evaluate PrivSpeech with interchanging privacy and utility setups with models for gender identification, emotion detection, and speaker verification. Our research extends the current understanding of utility, privacy, and energy consumption in the IoT landscape, offering new methodologies and privacy-preserving solutions. We expect to contribute to IoT systems designers, assisting them in making informed decisions to ensure privacy in an efficient and utility-preserving manner to IoT applications. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}