useckit 0.5.0a16

No project description provided

Useckit: An Open-Source Deep Learning Toolkit for Behavioral Biometrics

Useckit is an open-source Python toolkit designed for the development, evaluation, and deployment of deep learning-based user authentication systems. The toolkit bundles algorithms to evaluate behavioral biometrics for both user verification and identification tasks. It is offering a high-level API for quick experiments and a low-level API for custom model implementations. A full description can be found in our publication.

Overview

• Useckit bundles several deep learning paradigms for user authentication, including time-series classification, distance metric learning, and anomaly detection.
• Automatically computes key metrics like accuracy, F1-score, equal error rate (EER), receiver operating characteristic (ROC) curves, and more.
• Supports open-set and closed-set user identification as well as user verification.
• Customizable neural network models for advanced users.
• Extensible with custom datasets, preprocessing functions, and evaluation methods.

Usage

Installation

Use pip to install useckit and dependencies:

pip install useckit

Features

• Useckit offers a high-level API and a low-level API.
• The high-level API allows the quick application of predefined models that are grounded in literature.
  • We provide implementations, for example, for the approaches of Fawaz et al. (Time-Series Classification), Chen et al. (AutoEncoder-based authentication), and Schroff et al (Two-Stream Networks).
• The low-level API exists behind the facade of the high-level API and can also be used to use custom models within useckit.
• All results are automatically serialized to the filesystem.

Basic Usage (High-Level API)

To evaluate a dataset using default models:

import numpy as np
import useckit
from useckit.Evaluators import TSCEvaluator, DistanceLearningEvaluator

# Prepare dataset
x_train, y_train = np.array([...]), np.array([...])
x_test, y_test = np.array([...]), np.array([...])

dataset = useckit.Dataset(
    trainset_data=x_train,
    trainset_labels=y_train,
    testset_enrollment_data=x_train,
    testset_enrollment_labels=y_train,
    testset_matching_data=x_test,
    testset_matching_labels=y_test
)

# E.g., run time series classification evaluator
tse = TSCEvaluator(dataset, epochs=1000, verbose=False)
tse.evaluate()

# E.g., distance learning evaluator
dle = DistanceLearningEvaluator(dataset, epochs=1000, verbose=False)
dle.evaluate()

A comprehensible example can also be found in examples/useckit-high-level.ipynb.

Advanced Usage (Low-Level API)

You can customize models or extend Useckit with your own deep learning architectures. Examples and documentation are provided in examples/useckit-low-level.ipynb.

API Documentation

A sphinx-based documentation is work in progress and soon to be released. For now, we recommend checking the comprehensive examples and tests.

Features

Evaluation Methods

Useckit supports several evaluation paradigms, including:

• Verification Mode: For user verification tasks where a claim of identity is verified against a reference sample.
• Closed-Set Identification: For identifying users from a predefined set of identities.
• Open-Set Identification: Identifies known users and rejects unknown users.

Models

Useckit provides the following model architectures:

• Time Series Classification: FCN, Inception, ResNet, Encoder, TWIESN, MCDCNN, MLP, CNN (valid/same padding) padding, MCNN, t-leNet from Fawaz et al.
• Distance Learning: Two-stream networks with contrastive and triplet loss from Schroff et al.
• Outlier Detection: AutoEncoders from Chen et al.

Dataset Preparation

Datasets need to be provided in the following format:

• Train Set: Data and labels for model training.
• Validation Set: (Optional) Data and labels for model validation.
• Test Enrollment Set: Data and labels for system enrollment during testing.
• Test Matching Set: Data and labels for matching during testing.

Useckit supports k-fold cross-validation if your dataset is small.

Preprocessing Functions

The toolkit provides several preprocessing methods, including:

• Window Slicing: Sliding window for time-series data augmentation.
• Majority Voting: Aggregates predictions from window-sliced data.
• Normalization: Ensures data is normalized between [-1, +1].
• Data Checks: Verifies data integrity and format.

Evaluation Metrics

Useckit automatically computes the following key metrics:

• Accuracy
• Precision, Recall, F1-score
• Equal Error Rate (EER)
• Receiver Operating Characteristic (ROC) Curve
• Area Under the ROC Curve (AUC/AUROC)
• Confusion Matrix

Metrics are saved in both human-readable text format and machine-readable JSON format.

Authors

• Jonathan Liebers, University of Duisburg-Essen ([e-mail])
• Tristan Kley, University of Duisburg-Essen ([e-mail])
• Carina Liebers, University of Duisburg-Essen ([e-mail])
• Uwe Gruenefeld, University of Duisburg-Essen ([e-mail])
• Stefan Schneegass, University of Duisburg-Essen ([e-mail])

License

See LICENSE.