Gait recognition from inertial measurement unit (IMU) signals has emerged as a practical biometric modality for continuous and transparent user authentication. When deployed in practice, such systems must enroll new users over time through continual learning, i.e. fine-tuning the existing model on a new set of users to be enrolled. This operation introduces two intertwined challenges: catastrophic forgetting and privacy leakage. This thesis investigates these two problems jointly, using the Code Division Modulation Layer (CDML) framework as a starting point. The thesis extends this baseline in several directions and evaluates the resulting methods on three publicly available IMU datasets: WU Gait, UCI HAR and WISDM. On the learning side, a knowledge distillation strategy (CDML+KD), a Lipschitz constraint approach (LiDER) and a wavelet-based generative replay method (WGR-CDML) are proposed to reduce forgetting without storing raw biometric data. On the architectural side, Low-Rank Adaptation (LoRA) adapters are introduced as a parameter-efficient continual learning mechanism. A state-space model backbone based on the Mamba architecture (GaitMamba) is also evaluated as an alternative to the convolutional baseline. Privacy is evaluated against four attack families: Membership Inference Attacks, Identity Inference Attacks, Feature Space Inference and Backdoor Attacks. Finally, a set of cross-dataset experiments is presented, exploring transfer learning between datasets, adapter warm-starting and joint multi- dataset classification, to assess the generalisability of the proposed framework beyond the single-dataset setting
Gait recognition from inertial measurement unit (IMU) signals has emerged as a practical biometric modality for continuous and transparent user authentication. When deployed in practice, such systems must enroll new users over time through continual learning, i.e. fine-tuning the existing model on a new set of users to be enrolled. This operation introduces two intertwined challenges: catastrophic forgetting and privacy leakage. This thesis investigates these two problems jointly, using the Code Division Modulation Layer (CDML) framework as a starting point. The thesis extends this baseline in several directions and evaluates the resulting methods on three publicly available IMU datasets: WU Gait, UCI HAR and WISDM. On the learning side, a knowledge distillation strategy (CDML+KD), a Lipschitz constraint approach (LiDER) and a wavelet-based generative replay method (WGR-CDML) are proposed to reduce forgetting without storing raw biometric data. On the architectural side, Low-Rank Adaptation (LoRA) adapters are introduced as a parameter-efficient continual learning mechanism. A state-space model backbone based on the Mamba architecture (GaitMamba) is also evaluated as an alternative to the convolutional baseline. Privacy is evaluated against four attack families: Membership Inference Attacks, Identity Inference Attacks, Feature Space Inference and Backdoor Attacks. Finally, a set of cross-dataset experiments is presented, exploring transfer learning between datasets, adapter warm-starting and joint multi- dataset classification, to assess the generalisability of the proposed framework beyond the single-dataset setting
Privacy Attacks and Countermeasures in Biometric Gait Identification
DANIELI, TOMMASO
2025/2026
Abstract
Gait recognition from inertial measurement unit (IMU) signals has emerged as a practical biometric modality for continuous and transparent user authentication. When deployed in practice, such systems must enroll new users over time through continual learning, i.e. fine-tuning the existing model on a new set of users to be enrolled. This operation introduces two intertwined challenges: catastrophic forgetting and privacy leakage. This thesis investigates these two problems jointly, using the Code Division Modulation Layer (CDML) framework as a starting point. The thesis extends this baseline in several directions and evaluates the resulting methods on three publicly available IMU datasets: WU Gait, UCI HAR and WISDM. On the learning side, a knowledge distillation strategy (CDML+KD), a Lipschitz constraint approach (LiDER) and a wavelet-based generative replay method (WGR-CDML) are proposed to reduce forgetting without storing raw biometric data. On the architectural side, Low-Rank Adaptation (LoRA) adapters are introduced as a parameter-efficient continual learning mechanism. A state-space model backbone based on the Mamba architecture (GaitMamba) is also evaluated as an alternative to the convolutional baseline. Privacy is evaluated against four attack families: Membership Inference Attacks, Identity Inference Attacks, Feature Space Inference and Backdoor Attacks. Finally, a set of cross-dataset experiments is presented, exploring transfer learning between datasets, adapter warm-starting and joint multi- dataset classification, to assess the generalisability of the proposed framework beyond the single-dataset setting| File | Dimensione | Formato | |
|---|---|---|---|
|
Danieli_Tommaso.pdf
Accesso riservato
Dimensione
4.17 MB
Formato
Adobe PDF
|
4.17 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115860