Security and Privacy Enhanced Gait Authentication with Random Representation Learning and Digital Lockers

Lam Tran, Thuc Dinh Nguyen, Hyunil Kim, Deokjai ChoiarXiv 2021 · arXiv

Abstract

Gait authentication offers a promising method for implicit biometric security, identifying users by their unique walking patterns via smartphone sensors. However, traditional gait-based systems often store biometric templates in plaintext, posing significant privacy risks as biometric traits cannot be changed if compromised. This paper presents a novel gait-based cryptosystem that transforms sensor data into a secure, 139-bit cryptographic key. By integrating deep learning with a novel "Irreversible Error Correct and Obfuscate" (IECO) scheme, the system ensures that biometric signatures remain private, revocable, and secure against reverse-engineering attacks.

Introduction

Unlike passwords or PINs, gait authentication is an "implicit" biometric; it operates in the background without requiring active user interaction. While convenient, this approach faces two primary hurdles: security and stability. Most existing systems store gait templates in a format that, if stolen, allows an attacker to permanently compromise a user's biometric identity. Furthermore, gait data is inherently noisy—affected by fatigue, footwear, or terrain—making it difficult to generate the stable, exact bit-strings required for standard cryptographic applications. This research addresses these challenges by proposing a framework that generates high-entropy, cancelable cryptographic keys from smartphone sensor data.

Methodology

The proposed framework converts raw sensor data into a cryptographic key through a four-stage process designed to maximize both security and reliability.

1. Deep Feature Extraction (DFE)

The system utilizes a dual-branch neural network to process gait data. A Convolutional Neural Network (CNN) is employed to capture spatial gait postures, while a Long Short-Term Memory (LSTM) network captures the temporal dynamics of the user's stride. This combination ensures a comprehensive representation of the walking pattern.

2. Random Representation Learning

To ensure the generated keys are secure, the features must not only be distinct but also appear random. The authors developed a custom loss function that combines triplet loss—to ensure inter-user separability—with an inner product constraint. This constraint forces the features of different users to be nearly orthogonal. Consequently, the resulting binary strings behave like random noise to anyone other than the authorized user, achieving a Hamming distance of approximately 0.5 between different individuals.

3. Revocable String Forming (RSF)

To address the issue of permanent biometric compromise, the system incorporates Random Projection (RP). By mapping features into a new space using a random matrix, the system makes the template "cancelable." If a user's account is compromised, the system issues a new random matrix, effectively revoking the old template and generating a new, unrelated cryptographic key from the same biometric data.

4. Irreversible Error Correct and Obfuscate (IECO)

The final stage involves the IECO scheme, which improves upon traditional Error Correct and Obfuscate (ECO) methods. It utilizes "Digital Lockers"—a form of symmetric encryption secure against weak keys—to bind a random secret key to the gait string. This mechanism ensures that even if the final cryptographic key is leaked, an attacker cannot reverse-engineer the user's original gait pattern, maintaining strict privacy.

Experimental Evaluation

The model was rigorously tested on two benchmark datasets to evaluate its performance in both controlled and realistic environments:

  • OU-ISIR: The largest public gait dataset (744 users), used to test inter-user randomness across a large population.
  • whuGAIT: A dataset of 118 users recorded under "in the wild" conditions using smartphones, providing a test for real-world applicability.

Key Results

  • Security Strength: The system successfully generated keys of 139 bits and 147 bits, meeting and exceeding the 128-bit standard required for AES encryption.
  • Authentication Accuracy: On the OU-ISIR dataset, the model achieved a 0% False Acceptance Rate (FAR), ensuring no unauthorized users were granted access. The False Rejection Rate (FRR) stood at a manageable 4.167%.
  • Efficiency: The system requires only 5 gait cycles (approximately 5 seconds of walking) to authenticate. This is a significant improvement over previous gait-based cryptosystems that often required between 16 and 26 cycles.
  • Privacy Guarantees: Formal analysis confirmed that the system fulfills the requirements for unlinkability (keys from the same person appear unrelated) and irreversibility.

Discussion and Limitations

While the system demonstrates high security and accuracy, certain limitations remain. The performance metrics rely on "ideal" Digital Lockers; in practical deployments, the security parameter involves a trade-off between computational overhead and the certainty of key detection. Additionally, while the whuGAIT dataset covers realistic scenarios, further research is needed to determine how extreme environmental changes—such as walking on slippery surfaces or carrying heavy loads—might affect the False Rejection Rate. Finally, future work should explore cross-device generalization to ensure consistent performance across different smartphone hardware and sensor calibrations.

Conclusion

This research presents a significant step forward in biometric security. By combining deep representation learning with digital lockers and random projections, the authors have created a gait-authentication system that is not only accurate and fast but also provides the privacy and revocability necessary for modern cryptographic applications. This framework effectively bridges the gap between the inherent variability of human movement and the rigid requirements of digital security.

Summary updated

BibTeX
@misc{tran2021security,
  author = {Lam Tran and Thuc Dinh Nguyen and Hyunil Kim and Deokjai Choi},
  title = {{Security and Privacy Enhanced Gait Authentication with Random Representation Learning and Digital Lockers}},
  year = {2021},
  eprint = {2108.02400},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.02400},
}
Figure