Privacy-preserving human mobility and activity modelling
File(s)
Author(s)
Zhan, Yuting
Type
Thesis
Abstract
The exponential proliferation of digital trends and worldwide responses to the COVID-19 pandemic thrust the world into digitalization and interconnectedness, pushing increasingly new technologies/devices/applications into the market. More and more intimate data of users are collected for positive analysis purposes of improving living well-being but shared with/without the user's consent, emphasizing the importance of making human mobility and activity models inclusive, private, and fair. In this thesis, I develop and implement advanced methods/algorithms to model human mobility and activity in terms of temporal-context dynamics, multi-occupancy impacts, privacy protection, and fair analysis.
The following research questions have been thoroughly investigated: i) whether the temporal information integrated into the deep learning networks can improve the prediction accuracy in both predicting the next activity and its timing; ii) how is the trade-off between cost and performance when optimizing the sensor network for multiple-occupancy smart homes; iii) whether the malicious purposes such as user re-identification in human mobility modelling could be mitigated by adversarial learning; iv) whether the fairness implications of mobility models and whether privacy-preserving techniques perform equally for different groups of users.
To answer these research questions, I develop different architectures to model human activity and mobility. I first clarify the temporal-context dynamics in human activity modelling and achieve better prediction accuracy by appropriately using the temporal information. I then design a framework MoSen to simulate the interaction dynamics among residents and intelligent environments and generate an effective sensor network strategy. To relieve users' privacy concerns, I design Mo-PAE and show that the privacy of mobility traces attains decent protection at the marginal utility cost. Last but not least, I investigate the relations between fairness and privacy and conclude that while the privacy-aware model guarantees group fairness, it violates the individual fairness criteria.
The following research questions have been thoroughly investigated: i) whether the temporal information integrated into the deep learning networks can improve the prediction accuracy in both predicting the next activity and its timing; ii) how is the trade-off between cost and performance when optimizing the sensor network for multiple-occupancy smart homes; iii) whether the malicious purposes such as user re-identification in human mobility modelling could be mitigated by adversarial learning; iv) whether the fairness implications of mobility models and whether privacy-preserving techniques perform equally for different groups of users.
To answer these research questions, I develop different architectures to model human activity and mobility. I first clarify the temporal-context dynamics in human activity modelling and achieve better prediction accuracy by appropriately using the temporal information. I then design a framework MoSen to simulate the interaction dynamics among residents and intelligent environments and generate an effective sensor network strategy. To relieve users' privacy concerns, I design Mo-PAE and show that the privacy of mobility traces attains decent protection at the marginal utility cost. Last but not least, I investigate the relations between fairness and privacy and conclude that while the privacy-aware model guarantees group fairness, it violates the individual fairness criteria.
Version
Open Access
Date Issued
2022-07
Date Awarded
2022-12
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Haddadi, Hamed
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)