Supervised and Unsupervised Alignments for Spoofing Behavioral Biometrics

Thomas Thebaud, Gaël Le Lan, Anthony LarcherarXiv 2024 · DOI · arXiv

Abstract

Behavioral biometrics like voice and handwriting are increasingly used for authentication, often stored as high-dimensional embeddings. This paper investigates the vulnerability of these embeddings to template reconstruction attacks (TRAs) in a realistic "Architecture-Only" scenario. By employing supervised and unsupervised alignment techniques—specifically Wasserstein Procrustes—attackers can map stolen embeddings into a known space to generate high-quality spoofs. Our results demonstrate alarming success rates, with spoofing False Acceptation Rates (sFAR) reaching 54.64% for handwriting and 94.40% for speech, highlighting a critical privacy risk in biometric systems.

Introduction: The Vulnerability of Biometric Embeddings

In modern security architectures, biometric traits are converted into mathematical vectors known as embeddings. These embeddings are designed to be unique and robust, but their immutability poses a significant risk: if stolen, they cannot be "reset" like a password. Traditional Template Reconstruction Attacks (TRAs) often assume the attacker has access to the victim's specific encoder. This research addresses a more challenging and realistic "Architecture-Only" scenario, where the attacker knows the model type (e.g., ResNet34) but lacks access to the specific weights or training data. The core challenge lies in bridging the gap between the attacker's "proxy" embedding space and the victim's "target" space.

Methodology: Alignment and Reconstruction

The proposed attack pipeline utilizes alignment algorithms to translate vectors between two different, inaccessible coordinate systems.

Reconstruction Decoders

To turn embeddings back into raw behavioral data, the authors developed specialized decoders:

  • Handwriting: A novel LSTM-MDN (Mixture Density Network) decoder was introduced. Unlike standard LSTMs, the MDN predicts a probability distribution for subsequent points, effectively capturing the fluid and sharp-angled nature of human handwriting.
  • Speech: The system employs AutoVC, a many-to-many voice conversion model, modified with a specific spoofing reconstruction loss to optimize for identity preservation.

Alignment Techniques

Since the victim's and attacker's encoders produce distinct vector spaces, alignment is necessary.

  • Procrustes Analysis (Supervised): Finds the optimal rotation to map one set of points onto another using paired data.
  • Wasserstein Procrustes (Unsupervised): A key innovation that aligns two distributions without requiring paired samples. It uses stochastic optimization to "guess" the rotation based on the global geometric structure of the data clusters.
The attack pipeline using alignment matrices to bridge the gap between source and target embedding spaces.

Experimental Setup

The study utilized several diverse datasets to validate the attack:

  • Handwriting: Data from eBioDigit, MobileDigit, and a private dataset from Orange Innovation, consisting of digits drawn on touchscreens.
  • Speech: VoxCeleb1 and VoxCeleb2 were used for training encoders, while VCTK provided multi-speaker reading speech for training the voice conversion decoders.

Key Results and Insights

The experiments revealed a fundamental "geometric consistency" in biometric identity. Even when models are trained independently on different data, the relative distances between users in the embedding space remain similar. This intrinsic "shape" of identity allows unsupervised alignment to be devastatingly effective.

Examples of reconstructed handwriting digits (9, 3, 2, and 6) generated from embeddings.

Attack Performance

  • Handwriting: Using unsupervised Wasserstein Procrustes alignment, the attack achieved a Spoofing False Acceptation Rate (sFAR) of 54.64%, significantly higher than the baseline system error rate of 12.72%.
  • Speech: The results were even more potent, with unsupervised alignment reaching an sFAR of 94.40%. Remarkably, even with no alignment (Identity alignment), the system was spoofed 81.52% of the time, indicating high transferability of speech embeddings across models of the same architecture.

Conclusion and Future Directions

This work demonstrates that dynamic behavioral biometrics are as vulnerable to reconstruction as static images. The success of unsupervised alignment suggests that protecting the "architecture" alone is insufficient. Future research should investigate "cross-architecture" attacks and non-linear alignment methods. To mitigate these risks, the authors recommend implementing defensive measures such as bio-hashing, which shuffles the embedding space in a user-dependent manner to ensure privacy and compliance with regulations like GDPR.

Summary updated

Abstract

Biometric recognition systems are security systems based on intrinsic properties of their users, usually encoded in high dimension representations called embeddings, which potential theft would represent a greater threat than a temporary password or a replaceable key. To study the threat of embedding theft, we perform spoofing attacks on two behavioral biometric systems (an automatic speaker verification system and a handwritten digit analysis system) using a set of alignment techniques. Biometric recognition systems based on embeddings work in two phases: enrollment - where embeddings are collected and stored - then authentication - when new embeddings are compared to the stored ones -.The threat of stolen enrollment embeddings has been explored by the template reconstruction attack literature: reconstructing the original data to spoof an authentication system is doable with black-box access to their encoder. In this document, we explore the options available to perform template reconstruction attacks without any access to the encoder. To perform those attacks, we suppose general rules over the distribution of embeddings across encoders and use supervised and unsupervised algorithms to align an unlabeled set of embeddings with a set from a known encoder. The use of an alignment algorithm from the unsupervised translation literature gives promising results on spoofing two behavioral biometric systems.
BibTeX
@misc{thebaud2024supervised,
  author = {Thomas Thebaud and Gaël Le Lan and Anthony Larcher},
  title = {{Supervised and Unsupervised Alignments for Spoofing Behavioral Biometrics}},
  year = {2024},
  eprint = {2408.08918},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.08918},
}
Figure