Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PEaRL: Personalized Privacy of Human-Centric Systems using Early-Exit Reinforcement Learning

Mojtaba Taherisadr, Salma Elmalaki

TL;DR

The empirical results demonstrate PEaRL's capability to provide a personalized tradeoff between user privacy and application utility, adapting effectively to individual user preferences.

Abstract

In the evolving landscape of human-centric systems, personalized privacy solutions are becoming increasingly crucial due to the dynamic nature of human interactions. Traditional static privacy models often fail to meet the diverse and changing privacy needs of users. This paper introduces PEaRL, a system designed to enhance privacy preservation by tailoring its approach to individual behavioral patterns and preferences. While incorporating reinforcement learning (RL) for its adaptability, PEaRL primarily focuses on employing an early-exit strategy that dynamically balances privacy protection and system utility. This approach addresses the challenges posed by the variability and evolution of human behavior, which static privacy models struggle to handle effectively. We evaluate PEaRL in two distinct contexts: Smart Home environments and Virtual Reality (VR) Smart Classrooms. The empirical results demonstrate PEaRL's capability to provide a personalized tradeoff between user privacy and application utility, adapting effectively to individual user preferences. On average, across both systems, PEaRL enhances privacy protection by 31%, with a corresponding utility reduction of 24%.

PEaRL: Personalized Privacy of Human-Centric Systems using Early-Exit Reinforcement Learning

TL;DR

The empirical results demonstrate PEaRL's capability to provide a personalized tradeoff between user privacy and application utility, adapting effectively to individual user preferences.

Abstract

In the evolving landscape of human-centric systems, personalized privacy solutions are becoming increasingly crucial due to the dynamic nature of human interactions. Traditional static privacy models often fail to meet the diverse and changing privacy needs of users. This paper introduces PEaRL, a system designed to enhance privacy preservation by tailoring its approach to individual behavioral patterns and preferences. While incorporating reinforcement learning (RL) for its adaptability, PEaRL primarily focuses on employing an early-exit strategy that dynamically balances privacy protection and system utility. This approach addresses the challenges posed by the variability and evolution of human behavior, which static privacy models struggle to handle effectively. We evaluate PEaRL in two distinct contexts: Smart Home environments and Virtual Reality (VR) Smart Classrooms. The empirical results demonstrate PEaRL's capability to provide a personalized tradeoff between user privacy and application utility, adapting effectively to individual user preferences. On average, across both systems, PEaRL enhances privacy protection by 31%, with a corresponding utility reduction of 24%.
Paper Structure (30 sections, 2 equations, 16 figures, 3 tables, 3 algorithms)

This paper contains 30 sections, 2 equations, 16 figures, 3 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Privacy in machine learning
  4. Privacy-preserving RL
  5. Early-Exit in human-centric IoT
  6. Threat model in human-centric IoT Systems
  7. Privacy Leak Metric
  8. PEaRL algorithm
  9. Training DQN for EE with privacy and utility budgets
  10. PEaRL Inference
  11. Incooperating Human Variability
  12. Application 1: Thermal House Model
  13. Human-Centric System Design
  14. DRL Design
  15. Evaluation
  16. ...and 15 more sections

Figures (16)

  • Figure 1: Threat Model: A trusted edge device running a DQN shares recommended actions with an "honest-but-curious" cloud. The cloud, leveraging domain knowledge, may infer sensitive information from the action sequence, compromising user privacy.
  • Figure 2: Structure of the PEaRL algorithm. Each layer is followed by an exit branch and two confidence paths associated with the utility and privacy. Branches and confidence paths are trained independently.
  • Figure 3: PHASE 1 of the training flow of the PEaRL algorithm. An exit branch follows each layer. The layer and associated branch are trained in each step, and the next layer and its branch are added in the following step. Training of the new layer and its branch happens while the parameters of the previous layers and branches are frozen.
  • Figure 4: PHASE 2 of the training flow of the PEaRL algorithm. Privacy and utility confidence path training for each exit branch of the DQN. The network has two separate replay buffers. The privacy buffer contains the action, state, and privacy labels for all branches. The utility buffer includes the action, state, and utility labels of the branches.
  • Figure 5: Human activity profile for three humans.
  • ...and 11 more figures