[Paper Review] GAN pretraining for deep convolutional autoencoders applied to Software-based Fingerprint Presentation Attack Detection

Tobias Rohrer, Jascha Kolberg|arXiv (Cornell University)|May 21, 2021

Biometric Identification and Security4 citations

TL;DR

This paper proposes a novel transfer learning approach that uses a Wasserstein GAN to pretrain a deep convolutional autoencoder for one-class fingerprint presentation attack detection (PAD), achieving a 16.79% average classification error rate using only 1,122 bona fide fingerprint samples and no attack samples. The method enables robust feature learning from limited real data and generates synthetic fingerprint patches via the GAN for data augmentation and model pretraining.

ABSTRACT

The need for reliable systems to determine fingerprint presentation attacks grows with the rising use of the fingerprint for authentication. This work presents a new approach to single-class classification for software-based fingerprint presentation attach detection. The described method utilizes a Wasserstein GAN to apply transfer learning to a deep convolutional autoencoder. By doing so, the autoencoder could be pretrained and finetuned on the LivDet2021 Dermalog sensor dataset with only 1122 bona fide training samples. Without making use of any presentation attack samples, the model could archive an average classification error rate of 16.79%. The Wasserstein GAN implemented to pretrain the autoencoders weights can further be used to generate realistic-looking artificial fingerprint patches. Extensive testing of different autoencoder architectures and hyperparameters led to coarse architectural guidelines as well as multiple implementations which can be utilized for future work.

Motivation & Objective

• To address the challenge of training reliable one-class fingerprint presentation attack detectors with minimal training data.
• To overcome the difficulty of training deep convolutional autoencoders from scratch due to sensitivity to hyperparameters and lack of architectural guidelines.
• To explore the feasibility of using GANs not only for data generation but also for pretraining autoencoder weights in a one-class classification setup.
• To provide reusable, open-source implementations of WGAN, DCGAN, autoencoders, and data preprocessing pipelines for future PAD research.

Proposed method

• A Wasserstein GAN (WGAN) is trained on 1,122 bona fide fingerprint images to learn a robust feature distribution of real fingerprints.
• The trained generator and discriminator of the WGAN are used to pretrain the encoder and decoder of a deep convolutional autoencoder via transfer learning.
• The autoencoder is fine-tuned on the same bona fide dataset using a reconstruction loss, with input patches processed via data augmentation.
• The discriminator of the WGAN is used as a classifier by applying it to non-overlapping fingerprint patches and averaging the outputs for final prediction.
• Batch normalization, strided convolutions (instead of pooling), and a fully convolutional bottleneck are used to stabilize and improve autoencoder training.
• A data preprocessing pipeline applies extensive augmentation (cropping, flipping, rotating, brightness adjustment) to enhance generalization on limited real data.

Experimental results

Research questions

• RQ1 Can a WGAN be effectively used to pretrain a deep convolutional autoencoder for one-class fingerprint PAD without any presentation attack samples?
• RQ2 How does GAN-based pretraining compare to training autoencoders from scratch in terms of reconstruction accuracy and classification performance?
• RQ3 What architectural and hyperparameter choices significantly improve the training stability and performance of deep convolutional autoencoders for fingerprint PAD?
• RQ4 Can the WGAN generator produce diverse, realistic synthetic fingerprint patches useful for data augmentation or model evaluation?
• RQ5 To what extent does data augmentation improve the generalization of a one-class classifier trained on a very small number of bona fide samples?

Key findings

• The proposed WGAN-pretrained autoencoder achieved an average classification error rate of 16.79% on the LivDet2021 Dermalog dataset using only 1,122 bona fide samples and no attack samples.
• Training the autoencoder from scratch without transfer learning resulted in a significantly worse performance, with an average classification error rate of 23% under the same evaluation protocol.
• The WGAN successfully generated a wide variety of realistic synthetic fingerprint patches, supporting smooth latent space interpolation and visual diversity.
• The use of batch normalization, strided convolutions, and a fully convolutional bottleneck significantly improved training convergence and model stability.
• The discriminator of the WGAN alone, when applied to patch-level classification, did not yield reliable results, indicating limitations in direct GAN-based classification for this task.
• The authors released TensorFlow implementations of the data pipeline, WGAN, DCGAN, multiple autoencoder architectures, and variational autoencoders for reuse in future research.