With the advancement of digital systems, biometric authentication methods, especially fingerprint recognition, have become an integral component of various security protocols. However, these systems remain susceptible to spoofing attacks where counterfeit fingerprints can be used maliciously. This research aims to address the challenge of fingerprint liveness detection by leveraging the capabilities of Convolutional Neural Networks (CNNs). Using the SocoFing dataset, we designed and implemented a CNN-based model, highlighting its architecture comprising multiple convolutional layers, pooling layers, and dense layers. The model was trained with the Adam optimizer and evaluated using metrics such as accuracy, False Acceptance Rate (FAR), and False Rejection Rate (FRR). Our results offer promising insights into the robustness of CNNs in detecting genuine versus spoofed fingerprints. Furthermore, this study discusses challenges faced during implementation, implications for real-world applications, and potential avenues for future research in the realm of biometric security.
Biometric identification [1] methods provide reliable protection. Biometric user authentication is a method that identifies a user and verifies their identity based on the measurement of their unique physiological traits or behavioral characteristics [2]. Physiological biometrics are fingerprint, face recognition, iris scan, hand geometry, and retina scan. Behavioral biometrics are voice recognition, gait, keystroke scanning, and signature scanning. Fingerprints and handprints are the most widely used biometric methods today. Many laptops are equipped with fingerprint scanners, and fingerprint readers on USB drives are also available. Biometric authentication is widely used and has great reliability: it saves the user from the difficult task of recovering passwords; biometric data are unique and simple; it is very difficult to reproduce biometric characteristics; biometric characteristics cannot be lost; fingerprint scanning [3] is small and inexpensive; eye scanning is accuracy in user identification. 1.1.
With the ever-increasing reliance on biometric
authentication in security systems [
Liveness detection ensures [
While numerous techniques and
methodologies exist for fingerprint liveness
detection [
The primary objectives of this study are:
• To design and train a CNN model for
fingerprint liveness detection [
In addition to this, the study will shed light on potential improvements and directions for future research.
The research is based on the Socofing dataset
[
• About Socofing: The Socofing dataset
was created by the BiDA-Lab
(Biométrica de América Laboratory) at
the Universidad de las Fuerzas Armadas
ESPE in Ecuador. It aims to provide a
robust benchmark for algorithms aiming
to discern between genuine and fake
fingerprints.
• Size and Volume: The dataset comprises
numerous fingerprint images, with a
varied distribution between genuine and
spoofed fingerprints [
SocoFing dataset examples are displayed (Fig. 1): Real Fingerprints in the first row, Slightly modified fingerprints in the second row, and Greatly modified fingerprints in the third row.
The primary model explored in this research is a CNN, known for its prowess in handling image data. The architecture comprises: 1. Input Layer: • Purpose: Accepts the raw pixel values of the image as input. • Shape: Corresponds to the shape of train_images[0], which is the resolution of the fingerprint images used for training. If you’re using RGB images, this would be (height, width, 3), and for grayscale, it would be (height, width, 1). 2. Convolutional Layer (Conv2D): • Purpose: To scan the input image with filters/kernels, helping the model learn local patterns. • Number of Filters: 32. • Filter Size: 3×3. • Activation Function: ReLU (Rectified Linear Unit), which introduces nonlinearity to the model, enabling it to learn from the error and make adjustments, which is essential for learning complex patterns. 3. Max Pooling Layer (MaxPooling2D): • Purpose: To downsample the spatial dimensions of the output volume. It’s used to reduce the computational complexity, allowing the network to focus on the more relevant patterns/features. • Pool Size: 2×2. 4. Convolutional Layer (Conv2D): • Purpose: Another convolution operation to extract higher-level features. • Number of Filters: 64. • Filter Size: 3×3. • Activation Function: ReLU. 5. Max Pooling Layer (MaxPooling2D): • Purpose: Again, downsampling the spatial dimensions. • Pool Size: 2×2. 6. Flattening Layer (Flatten): • Purpose: As its name suggests, this layer flattens the output of the previous layers into a single long vector. This is necessary because fully connected layers (dense layers) expect a 1D input vector. 7. Fully Connected Layer (Dense): • Purpose: It interprets the features and patterns learned by previous layers. • Units (Neurons): 128. • Activation Function: ReLU. 8. Output Layer (Dense): • Purpose: To output a probability distribution over the classes (Genuine and Spoof in your case). • Units (Neurons): 2, corresponding to the two classes: Genuine and Spoof.
Activation Function: Softmax. This activation function returns the probability distribution over the classes, meaning each neuron will output a value between 0 and 1, representing the likelihood of the input image belonging to its respective class (Fig. 2).
Optimizer: The Adam optimizer is an adaptive
learning rate optimization algorithm [
Loss Function: Sparse Categorical Cross EntropyThis loss function is used for multiclass classification problems where the labels are integers (as opposed to one-hot encoded vectors). It computes the cross-entropy loss between true labels and predicted labels. 2.3.
To measure the performance of the models, the
following metrics were employed:
• Accuracy: This represents the
proportion of correctly classified
fingerprints out of the total fingerprints
[
For this research, we aimed to evaluate the effectiveness of a Convolutional Neural Network (CNN) in distinguishing between genuine and spoofed fingerprints using the Socofing dataset: • Data Preprocessing: Before feeding the data to the CNN, each fingerprint image underwent standard pre-processing steps, which included resizing to a consistent resolution, normalization to scale pixel values between 0 and 1, and data augmentation (like random rotations and flips) to enhance the model’s robustness. • Model Initialization: The chosen CNN architecture comprised multiple convolutional, pooling, and dense layers. The model was initialized with random weights. • Training Settings: Training was performed using a batch size of 2764 and for 10 epochs. We employed a split of 80% for training and 20% for validation from the dataset. The Adam optimizer was used with a learning rate of 0.001 and a categorical cross-entropy loss function. Throughout the training process, we monitored both the training and validation loss and accuracy (Table 1). The convergence of training and validation curves indicated minimal overfitting, and the model generalizes well to unseen data (Fig. 3).
Upon completion of the training process, the model was evaluated on a test dataset to measure its performance: • Accuracy: 98.964%. • False Acceptance Rate (FAR): 0.215%. • False Rejection Rate (FRR): 7.251%.
The model showcased a high accuracy rate, indicating its proficiency in distinguishing between genuine and spoofed fingerprints. However, the FRR suggests a need for further optimizations, as genuine fingerprints were occasionally misclassified (Fig. 4). While the model showed a high rate of true positives and true negatives, indicating correct classifications, there were instances of both false positives and false negatives, which emphasize areas for potential model refinement (Fig. 5).
The CNN’s performance on the Socofing dataset offered several valuable insights: • High Accuracy: The CNN demonstrated a strong capability in distinguishing genuine from spoofed fingerprints with an accuracy nearing 99%. This underscores the model’s potential as a robust tool for fingerprint liveness detection. • Potential for Optimization: Despite the impressive accuracy, the FRR of 7.251% highlights an area that requires attention. In practical scenarios, genuine users could face challenges in authentication due to such misclassifications. • Confusion Matrix Observations: The low number of false positives suggests that the model rarely misidentifies a spoof as a genuine fingerprint. However, the presence of false negatives indicates room for enhancing the model's sensitivity. 4.2.
During the research, several challenges were faced: • Data Imbalance: The Socofing dataset might contain an imbalanced distribution between genuine and spoofed fingerprints, which can influence model training. • Data Variability: The varying quality, orientation, and characteristics of fingerprints in the dataset can pose challenges to model consistency. • Computational Constraints: Training deep learning models requires significant computational resources, which might limit the extent of hyperparameter tuning and experimentation.
From a practical standpoint, the model’s
results emphasize its utility in security systems
requiring fingerprint authentication [
The research embarked on an exploration of the CNN-based Model’s efficacy in fingerprint liveness detection using the Socofing dataset. It displayed a commendable ability, with an accuracy close to 99%. However, certain metrics revealed avenues for optimization.
Apart from the theoretical variety of
possible biometric methods used in practice
[
Based on the findings: • Data Augmentation: Increase the use of data augmentation techniques to enhance the model’s ability to recognize diverse fingerprint variations. • Model Refinement: Dive deeper into model architecture, considering additions or modifications, to address the elevated FRR.
Continuous Training: Keep the model updated with new data to ensure it stays relevant against emerging spoofing techniques. 5.3.
Future research could focus on: • Hybrid Models: Combining the strengths of different architectures, like integrating features from RNNs or Autoencoders with CNNs, might offer improved performance. • Transfer Learning: Utilize pre-trained models on larger datasets and finetune them for fingerprint liveness detection. • Advanced-Data Augmentation: Techniques like GANs can generate synthetic fingerprints to expand the dataset and potentially improve model robustness.