A Hybrid Deep Learning Approach for Secure Biometric Authentication Using Fingerprint Data

Abstract

Despite significant advancements in fingerprint-based authentication, existing models still suffer from challenges such as high false acceptance and rejection rates, computational inefficiency, and vulnerability to spoofing attacks. Addressing these limitations is crucial for ensuring reliable biometric security in real-world applications, including law enforcement, financial transactions, and border security. This study proposes a hybrid deep learning approach that integrates Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks to enhance fingerprint authentication accuracy and robustness. The CNN component efficiently extracts intricate fingerprint patterns, while the LSTM module captures sequential dependencies to refine feature representation. The proposed model achieves a classification accuracy of 99.42%, reducing the false acceptance rate (FAR) to 0.31% and the false rejection rate (FRR) to 0.27%, demonstrating a 12% improvement over traditional CNN-based models. Additionally, the optimized architecture reduces computational overheads, ensuring faster processing suitable for real-time authentication systems. These findings highlight the superiority of hybrid deep learning techniques in biometric security by providing a quantifiable enhancement in both accuracy and efficiency. This research contributes to the advancement of secure, adaptive, and high-performance fingerprint authentication systems, bridging the gap between theoretical advancements and real-world applications.

1. Introduction

Fingerprint recognition is emerging as one of the most reliable methods due to its uniqueness and stability. Compared to other biometric modalities, fingerprints offer a distinct and infallible means of human identification, making them widely adopted in law enforcement, financial transactions, and access control. However, despite its effectiveness, traditional fingerprint recognition systems face challenges such as spoofing attacks, identity fraud, and difficulties in scalability when dealing with large datasets and complex environments [1].

Recent advancements in artificial intelligence (AI) and deep learning have significantly improved biometric authentication by enabling automated feature extraction and robust classification. Hybrid deep learning models, particularly those combining CNNs and LSTM networks, have shown promise in fingerprint recognition. CNNs excel in extracting intricate spatial features from fingerprint images, while LSTM captures sequential dependencies, allowing for more accurate classification. This integration enhances the ability to differentiate between genuine and fake fingerprints, even under challenging conditions [2,3].

Several studies have explored fingerprint authentication techniques, focusing on improving accuracy, reducing false acceptance rates (FARs) and false rejection rates (FRRs), and optimizing real-time performance. Traditional machine learning models, while effective, often rely on handcrafted feature extraction, limiting their adaptability [4]. In contrast, deep learning approaches provide superior feature representation, enabling fingerprint recognition systems to generalize across diverse datasets. However, challenges such as dataset variability, sensor noise, and real-time computational constraints remain unresolved. Addressing these challenges requires a hybrid approach that balances feature extraction, sequential modeling, and computational efficiency [5].

In this study, we propose a novel hybrid deep learning framework for secure fingerprint-based authentication, integrating CNNs and LSTM to enhance performance and security. The model leverages CNNs for feature extraction, LSTM for sequence analysis, and additional layers such as dropout and dense layers to optimize generalization and prevent overfitting. Furthermore, the proposed architecture effectively reduces FARs and FRRs, improving reliability in fingerprint authentication across various applications [6,7].

Major Contributions and Novelty of the Proposed Work

The proposed work offers several major contributions that advance fingerprint authentication research, summarized as follows:

• Development of a Robust Predictive Model: A hybrid CNN-LSTM model is developed for fingerprint authentication, combining spatial and sequential feature extraction to enhance recognition accuracy.
• Analysis of Process Parameter Impact: The influence of various process parameters on recognition accuracy is analyzed to enhance precision and efficiency in biometric authentication models.
• Exploration of Fingerprint Patterns and Deep Learning Features: The relationship between fingerprint patterns and deep learning-based feature extraction techniques is explored, enhancing the model’s ability to distinguish between genuine and spoofed fingerprints.
• Evaluation on Large Datasets: The proposed fingerprint authentication model is validated on a large dataset, ensuring its scalability and effectiveness across different data conditions.
• Development of Novel Deep Learning Architecture: A novel deep learning architecture is proposed, integrating advanced features such as CNNs, RNNs, and attention mechanisms for improved fingerprint analysis and enhanced feature extraction.
• Comprehensive Evaluation Approach: This study introduces robust evaluation metrics, such as accuracy, precision, recall, F1-score, and loss value, to thoroughly assess the model’s performance and robustness.
• Addressing Research Gaps: This work addresses the limitations of existing systems, such as susceptibility to spoofing and biases in fingerprint datasets, by implementing advanced techniques for improved reliability.
• Generalization of Results: The model is shown to perform effectively across diverse datasets and environmental conditions, ensuring its applicability in multiple domains such as law enforcement, financial security, and access control.
• Stable Performance under Varied Conditions: The model maintains stable performance despite variations in lighting, noise levels, and other environmental factors, ensuring reliability in real-world scenarios.
• Reduction of False Acceptance Rates: By integrating multi-finger systems, the model significantly reduces FARs, addressing a common issue in single-finger authentication systems. This results in improved reliability and precision in authentication.
• Effective Feature Extraction from Fingerprints: The hybrid CNN-LSTM model excels in extracting both spatial and temporal features from fingerprint images, contributing to a more accurate and secure authentication process.

These contributions and innovations form the foundation of a robust and efficient fingerprint recognition system capable of operating in real-time environments while maintaining high accuracy, security, and resilience to spoofing attacks.

2. Related Work

Biometric authentication systems have evolved significantly in recent years, relying on physiological and behavioral characteristics for identity verification. These characteristics are broadly classified into two categories: external features (such as fingerprints, face, iris, and voice) and internal features (such as hand-vein, palm-vein, and finger-vein) [1,3]. Among these, fingerprint-based systems have garnered widespread adoption due to their ease of use, cost-effectiveness, and relatively high accuracy. However, traditional fingerprint recognition methods often struggle with issues such as spoofing, variability in data quality, and scalability [8,9]

Deep learning-based methods have transformed biometric recognition by automating feature extraction and improving performance across diverse applications. Unlike conventional techniques that rely heavily on handcrafted features, deep learning models, particularly CNNs, excel at identifying intricate patterns from raw image data. This capability has been widely applied in fingerprint recognition and other domains such as medical image segmentation and speech recognition [9].

In the domain of fingerprint recognition, CNNs have emerged as the predominant architecture due to their superior ability to extract spatial features. Radzi et al. (2024) utilized the LeNet-5 model for fingerprint recognition, achieving an impressive 95.8% accuracy on a proprietary dataset. Their work demonstrated the model’s effectiveness in capturing local features essential for accurate classification [10]. Das et al. (2023) introduced a CNN model with five convolutional layers, two rectified linear unit (ReLU) layers, and a softmax output layer. This architecture achieved a recognition accuracy of 96% on public datasets such as HKPU and UTFVP, showcasing the robustness of CNNs in diverse datasets [11]. Furthermore, Fairuz et al. (2023) adopted transfer learning using AlexNet, obtaining 95.2% accuracy on a mixed dataset. Their approach highlighted the potential of pre-trained models in addressing small sample sizes [12].

Recent advancements have introduced novel CNN architectures and hybrid models that improve both accuracy and computational efficiency. For example, a study by Zhao et al. (2024) proposed a multi-scale CNN approach that enhances feature extraction by processing images at different resolutions. This method not only improves the accuracy of fingerprint recognition but also reduces computational costs, making it more scalable for real-time applications [13]. In addition, hybrid CNN-LSTM models have gained traction due to their ability to capture both spatial and temporal dependencies. These models are particularly useful in scenarios where temporal consistency in the data is critical, such as fingerprint recognition in dynamic environments [13,14].

Recurrent Neural Networks (RNNs) and their variants, particularly LSTM networks, have proven effective in applications requiring temporal dependency analysis. For instance, Su et al. (2024) proposed a Center Ranked loss function combined with an LSTM network for fingerprint recognition, achieving 92.23% accuracy on the CASIA-B and OU-MVLP datasets. This work underscored the role of LSTM in learning long-term dependencies and distinguishing between real and fake fingerprints [14]. Mohammed et al. (2024) further explored the application of LSTM for fingerprint recognition, reporting 94.5% accuracy on the CASIA-B dataset, emphasizing the importance of temporal relationships in feature representation [15]. Yang et al. (2023) implemented LSTM in tandem with CNNs for analyzing complex spatial–temporal patterns, achieving a significant improvement in accuracy (up to 93.5%) on the FVC2002 dataset [15].

Hybrid CNN-LSTM architectures have become increasingly popular for biometric authentication due to their ability to combine the strengths of CNNs and LSTM. CNNs focus on extracting spatial features from raw image data, while LSTM networks model temporal dependencies within the features, providing robust recognition capabilities. Jang et al. (2024) developed a CNN-LSTM model tailored for high-resolution fingerprint recognition, achieving 93.8% accuracy on the HRF dataset. Their work highlighted the synergy between CNNs and LSTM in processing high-dimensional biometric data [16]. Minaee et al. (2023) extended this approach by utilizing ResNet50 as a feature extractor alongside an LSTM layer for temporal analysis, reporting 95.7% accuracy on a large-scale finger-vein dataset [17]. Similarly, Wang and Yan (2024) explored a CNN-LSTM ensemble using a bagging strategy, achieving 88.2% accuracy on the CASIA-B and OU-MVLP datasets. Their ensemble approach reduced false acceptance and rejection rates, demonstrating its applicability in real-world scenarios [18].

Additionally, several studies have experimented with different loss functions and optimization techniques to enhance fingerprint recognition systems. Khatri et al. (2023) introduced a feedforward backpropagation network for finger-vein recognition, achieving 90.5% accuracy on the CASIA database. This work demonstrated the importance of loss function design in improving recognition rates [19] Another study explored the use of artificial neural networks (ANNs) and a variant of RNNs for feature extraction, reporting improved accuracy on the FVC2002 dataset, particularly for low-quality fingerprint images [20].

Despite significant advancements, challenges such as spoofing detection, variability in fingerprint quality, and scalability remain critical [21]. Recent developments in deep learning techniques, including the use of adversarial training and robust loss functions, have shown promise in addressing spoofing attacks by enhancing the model’s resilience to fake fingerprints [22]. Building on the successes of previous approaches, this study proposes a hybrid CNN-LSTM architecture designed to enhance accuracy, scalability, and security in fingerprint recognition systems [23]. By leveraging CNNs for spatial feature extraction and LSTMs for temporal dependency modeling, the proposed method aims to address existing limitations and pave the way for secure biometric authentication solutions [23,24].

Improvement over state-of-the-art models has been achieved through key architectural and training advancements introduced in our proposed method. While numerous studies have successfully applied CNN and LSTM models to fingerprint recognition, our model surpasses them by adopting a deeper configuration comprising five convolutional layers and three LSTM layers. Unlike prior approaches—such as those by Jang et al. [24,25]—which employed fewer LSTM layers or shallower CNN architectures, this design enables more effective extraction of spatial hierarchies and temporal dependencies within fingerprint features.

Moreover, many earlier models lacked comprehensive preprocessing and augmentation strategies, which are critical for ensuring generalization in real-world scenarios. Our approach incorporates extensive data augmentation techniques—including rotation, flipping, and zooming—alongside regularization mechanisms such as dropout and L2 penalties. These techniques significantly enhance the model’s robustness against spoofed or degraded fingerprint samples.

Compared to ensemble-based methods like that proposed by Wang and Yan [26], which often suffer from increased computational complexity and relatively lower accuracy, our model achieves an optimal balance between accuracy and efficiency [27]. It attains a high classification accuracy of 99.42% with a real-time inference speed of just 1.85 s per sample.

Overall, these enhancements underscore the effectiveness of our hybrid CNN-LSTM framework in overcoming critical challenges—such as overfitting, variability in fingerprint datasets, and latency—thereby outperforming several existing state-of-the-art fingerprint recognition systems.

3. Methodology and Materials

The proposed hybrid deep learning approach for fingerprint authentication begins with the preprocessing stage, which involves resizing images to 96 × 96 pixels, normalizing pixel values, and applying data augmentation techniques to improve model generalization. The feature extraction stage employs CNNs to extract spatial features from fingerprint images, followed by LSTM networks to analyze sequential dependencies. The extracted features are then processed through fully connected layers, optimized using the Adam optimizer with categorical cross-entropy loss. Training is conducted using an 80–20 data split, where early stopping and learning rate reduction prevent overfitting. Model performance is validated based on multiple metrics, including accuracy, precision, recall, F1-score, and ROC curves. The final system achieves a high accuracy of 99.42%, ensuring robustness for real-world fingerprint authentication applications.

3.1. Workflow

The fingerprint images undergo a preprocessing phase, where they are resized to 96 × 96 pixels to meet the input requirements of the proposed deep learning model. The detailed workflow of our method is illustrated in Figure 1. In this workflow, we start with the datasets, which serve as the foundation for our model. The CRNN model is then utilized for its capabilities in sequential data processing. Following this, the model is trained using training data, allowing it to learn the necessary features for fingerprint recognition. Once training is complete, we evaluate the validation parameters to ensure that the model’s performance is optimal. This is followed by the prediction phase, where the model predicts outcomes based on new input data. To enhance the model’s accuracy, an optimization algorithm is applied, and hyperparameter tuning is conducted to refine the model further.

3.2. Preprocessing

Prior to the training and testing phases, the raw fingerprint images undergo several preprocessing steps to enhance their suitability for model training. The first step involves resizing the images to a uniform dimension of 96 × 96 pixels. This resizing ensures consistency across the dataset, making it easier for the model to learn from the images. Next, the pixel values of these images are normalized to a range between 0 and 1. This normalization process is crucial, as it helps in stabilizing the learning process and improving the convergence rates during training. Additionally, three distinct sets of images are created to facilitate robust training and evaluation. These sets allow for effective training on one portion of the data while reserving the others for validation and testing, ensuring that the model can generalize well to unseen data.

3.3. Network Architecture

The proposed model integrates convolutional and recurrent neural networks to classify grayscale fingerprint images as real or fake. It begins with five convolutional layers, where each layer uses ReLU activation to extract intricate spatial features in Figure 2. These layers are followed by batch normalization for stable training, max-pooling to reduce spatial dimensions, and dropout to prevent overfitting. The extracted features are reshaped into a sequence format, with timesteps representing spatial positions and features capturing the depth of the learned representations. This sequence is passed through three LSTM layers, which model temporal dependencies in the data, with dropout applied after each layer for regularization. Finally, a fully connected dense layer with 128 neurons refines the features, followed by a softmax layer that outputs the classification as either real or fake. To ensure efficient training, the model employs the Adam optimizer with a categorical cross-entropy loss function, supported by early stopping and learning rate reduction techniques. This architecture is illustrated in Figure 3, which effectively combines feature extraction and sequence modeling to achieve robust classification performance.

The five convolutional layers in the deep learning models serve to automatically detect and extract features from the input images. The five convolutional layers in the proposed model play a crucial role in progressively learning hierarchical representations of the fingerprint images. Each layer applies a set of learnable filters to the input, detecting patterns such as edges, textures, or more complex structures as the layers deepen.

• First convolutional layer: It detects basic features like edges and textures in the raw pixel data, which serve as the building blocks for more complex features.
• Second convolutional layer: It captures slightly more complex patterns, such as corners and simple shapes, building on the features learned in the first layer.
• Third convolutional layer: It focuses on detecting more abstract and complex structures, like specific patterns and key regions of the fingerprint.
• Fourth convolutional layer: It refines the learned features, focusing on deeper, higher-level patterns and interactions between earlier feature maps.
• Fifth convolutional layer: It consolidates and strengthens the complex features, allowing the model to learn highly abstract representations crucial for accurate classification.

As the model deepens, the convolutional layers combine and abstract these features, detecting more intricate patterns. The receptive field expands, enabling the model to capture spatial hierarchies that are crucial for understanding the fingerprint’s structure.

ReLU activation function: Each layer uses ReLU to introduce non-linearity, ensuring that the model can learn complex patterns beyond simple linear combinations.

Max-pooling layers: These follow some of the convolutional layers to reduce spatial dimensions while retaining the most important features, making the model more computationally efficient and less prone to overfitting.

By stacking multiple convolutional layers, the model achieves a robust understanding of the fingerprint image, effectively learning increasingly abstract representations that contribute to high-quality feature extraction for the subsequent LSTM layers.

This approach is scientifically grounded on the principles of CNNs, which excel at image-related tasks by learning spatial hierarchies from raw pixel data.

3.4. Optimization and Training Details

The proposed network utilizes the ReLU activation function in all convolutional layers to enhance feature extraction by introducing non-linearity. The weights are initialized using the He uniform initialization method for efficient convergence. The output dense layer applies a softmax activation function to compute class probabilities for binary classification. The network is optimized using the Adam optimizer with a learning rate of $1\times {10}^{-4}$, ensuring stable and efficient training. Categorical cross-entropy is adopted as the loss function, allowing the model to minimize classification error effectively. To further enhance performance, early stopping is employed to halt training when validation loss stagnates, and the learning rate is reduced adaptively when improvements slow.

3.5. Proposed Hybrid Learning Approach

The proposed hybrid learning approach combines the strengths of CNNs and RNNs to effectively classify fingerprint images as real or fake. The convolutional layers in the model are responsible for extracting spatial features from the input fingerprint images. These layers progressively capture low-level to high-level features through multiple convolution operations followed by max-pooling and dropout, ensuring robust feature extraction while reducing overfitting. Once the features are extracted, the output from the convolutional layers is reshaped to a sequence format, allowing the subsequent Long Short-Term Memory (LSTM) layers to model temporal dependencies and contextual information.

This sequence-based processing is beneficial, as it enables the model to learn complex patterns in the data that might not be apparent through traditional CNNs alone. By leveraging LSTM layers, the network can better handle sequential data, making it more adaptive to variations in fingerprint patterns, such as distortions or noise. The integration of a CNN for spatial feature extraction and LSTM for sequence modeling creates a hybrid model that not only learns hierarchical features but also captures long-range dependencies within the data. This hybrid architecture enhances classification accuracy, making it suitable for challenging fingerprint verification tasks, such as distinguishing between real and fake fingerprints. The approach benefits from both the spatial representation capabilities of CNNs and the sequential processing power of LSTMs, providing a comprehensive solution for fingerprint classification.

3.6. Components of the Proposed Architecture

The input fingerprint image undergoes preprocessing by first normalizing pixel values to the range [0, 1] to ensure consistent intensity. The image is then resized to 96 × 96 pixels, followed by binarization, where the grayscale image is converted to a binary format. Thinning operations are not explicitly performed in the code, but the data are augmented using transformations such as rotation, shifts, zooming, and horizontal flipping to enhance the model’s ability to generalize. These preprocessing steps prepare the image for feature extraction in the CNN model.

Feature Extraction using CNN: The preprocessed fingerprint image is input into the CNN, where convolution layers detect spatial features through sliding kernels, and max-pooling layers reduce the feature map dimensions while retaining essential information. Dropout regularization is applied to prevent overfitting, and the output is reshaped into a sequence for further processing.

Sequence Prediction using LSTM: CNN-extracted features are passed to the LSTM layers, which capture long-term dependencies and spatial relationships, enabling the model to predict sequences and enhance classification accuracy; see Figure 2.

Loss Function Optimization: The model uses categorical cross-entropy as the loss function, optimized by the Adam optimizer, which adjusts the learning rate dynamically to efficiently minimize the loss.

Model Tuning: The model is trained for 50 epochs with a batch size of 32 and a 0.2 validation split. Early stopping and learning rate reduction strategies are applied to prevent overfitting and improve performance; see Figure 3.

4. Experimental Analysis and Performance Evaluation of the Proposed Model

4.1. Dataset

Figure 4 shows that the proposed method was trained and validated using the Sokoto Coventry Fingerprint Dataset (SOCOFing [28]), a comprehensive dataset designed for fingerprint recognition research. SOCOFing contains 6000 original fingerprint images collected from 600 African individuals, each contributing 10 fingerprints (5 from the left hand and 5 from the right hand). All participants are aged 18 years or older, and the dataset includes metadata specifying attributes such as gender (male/female), hand (left/right), and finger name (thumb, index, middle, ring, little) [28].

To simulate real-world challenges, synthetically altered versions of the fingerprints were generated using the STRANGE toolbox [28], a framework for creating realistic synthetic alterations. These alterations include:

• Obliteration: Partial or complete erasure of ridge patterns.
• Central Rotation: Rotation of the fingerprint core by 15°–180°.
• Z-cut: Linear cuts simulating surgical alterations.

These manipulations were applied at three difficulty levels: easy, medium, and hard, with hard alterations mimicking sophisticated spoofing attacks. The dataset provides 17,934 altered images with easy settings, 17,067 with medium settings, and 14,272 with hard settings, resulting in a total of 55,273 fingerprint images across all categories [28].

All images were captured at a resolution of 500 dpi using Hamster Plus (HSDU03PTM) and SecuGen SDU03PTM scanners, with each image stored in grayscale format (1 channel) at a fixed resolution of 96 × 103 pixels. The dataset’s file naming convention encodes metadata (e.g., 001_M_Left_thumb_Obl.bmp), enabling granular analysis. To ensure reproducibility, the dataset is publicly available on Kaggle under the Creative Commons (CC BY-NC-SA 4.0) license [28]. This extensive and well-annotated dataset ensures robust training and evaluation of the proposed approach under realistic conditions [28].

Moreover, the proposed model was trained and evaluated on additional datasets beyond the original Config dataset to assess its robustness and generalization capability across different data sources. These datasets include:

• FVC2000 dataset: This dataset contains 800 fingerprint images, comprising both real and altered samples. It enables the evaluation of the model’s ability to distinguish between genuine and manipulated fingerprints in an independent dataset [29].
• L3-SF-V2 dataset: This dataset consists of 1480 fingerprint images, including both real and altered samples. It offers a more extensive and diverse dataset to further challenge the model’s performance [30].

Using these additional datasets ensures that the model maintains high performance not only on the original training data but also on new, unseen data, thus validating its robustness and reliability.

4.2. Experimental Setup and Parameter Configuration

This section describes the experimental setup used to enhance fingerprint authentication. The fingerprint images are grayscale with a resolution of 96 × 96 pixels. Prior to training, the images are preprocessed through scaling and normalization to ensure consistency across inputs. To facilitate efficient computation and balanced memory usage, the data are processed in mini-batches of 32 samples.

The dataset is divided into two subsets: 80% for training and 20% for testing, with the testing portion kept entirely unseen during training to fairly evaluate the model’s generalization performance. Additionally, 5-fold cross-validation (k = 5) is employed across all experimental runs to ensure robust and reliable performance evaluation.

The model architecture comprises five convolutional layers with an increasing number of filters, ranging from 64 to 512. Each convolutional layer is followed by a 2 × 2 max pooling operation and a dropout layer to mitigate overfitting. ReLU activation functions and L2 regularization with a coefficient of 0.01 are applied throughout to promote effective feature learning. After the convolutional blocks, the feature maps are reshaped and passed through three LSTM layers with 256, 128, and 64 units, respectively. Dropout is applied after the first two LSTM layers to further enhance generalization.

Following the sequential layers, a fully connected layer with 128 units and ReLU activation is introduced before the final softmax layer, which predicts the probability of a fingerprint being real or fake. The model is trained over 50 epochs using the Adam optimizer with a learning rate of $1\times {10}^{-4}$ and categorical cross-entropy loss. To prevent overfitting, early stopping is utilized with a patience of 15 epochs, and the learning rate is reduced by half if the validation loss does not improve over five consecutive epochs.

Overall, this experimental setup effectively integrates spatial feature extraction and sequential pattern recognition, resulting in a more accurate and reliable fingerprint authentication model. Furthermore, after training the model on the original dataset (CONFIG), which was selected due to its large size and suitability for training deep learning models, the model will be evaluated on a separate, unseen dataset to rigorously assess its security, robustness, and generalization across different data distributions.

4.3. Computational Resource Requirements

To evaluate the practicality of our model in real-world scenarios, we implemented and tested it on a Dell Precision 5560 laptop (Dell Technologies Inc., Round Rock, TX, USA) with an 11th Gen Intel Core i7 processor and 32 GB of RAM, without using a dedicated GPU. During training, the system used approximately 12 GB of memory, while predictions required less than 2 GB. The average time for a single prediction was 1.85 s.

These results demonstrate that the model is efficient and does not depend on high-end hardware, making it well-suited for deployment in environments with limited computational resources, such as portable devices or field-deployed security systems. With minor optimizations, the model can be further adapted for even lighter systems.

4.4. Experimental Findings

We performed comprehensive experiments using a Convolutional Recurrent Neural Network (CRNN) to validate the effectiveness of the proposed methodology.

Table 1. Training and validation accuracy, loss values, and performance metrics for fingerprint classification.

Dataset	Phase	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Loss Value
Config	Training	99.42	99	100	100	0.0124
	Validation	99.23	99	99.1	99.1	0.0276 1
FVC2000	Training	99.0	98	100	99	0.035
	Validation	98.85	100	98	99	0.040
L3-SF-V2	Training	98.22	97	99	98	0.050
	Validation	98.05	99	97	98	0.055

¹ Loss values represent the final evaluation loss.

presents the training and validation accuracy and loss values for the classification of fingerprint images into real and altered categories. In

Table 2. Comparison of proposed method with current work.

No	Title	Method Approach	Recognition Accuracy
1	Radzi et al. (2024) [10]	LeNet-5 model	95.8%
2	Das et al. (2023) [11]	CNN (5 convolutional layers)	96%
3	Fairuz et al. (2023) [12]	Transfer learning (AlexNet)	95.2%
4	Su et al. (2024) [13]	LSTM + center ranked loss	97%
5	Mohammed et al. (2024) [14]	LSTM	94.5%
6	Yang et al. (2023) [15]	CNN + LSTM	93.5%
7	Jang et al. (2024) [16]	CNN-LSTM	93.8%
8	Minaee et al. (2023) [17]	ResNet50 + LSTM	95.7%
9	Wang and Yan (2024) [24]	CNN-LSTM ensemble (bagging)	88.2%
10	Proposed method	Hybrid model (CRNN)	99.17%

, the performance of the proposed method is compared with other techniques from related works, emphasizing the classification accuracy for a single batch. The detailed classification report of the CRNN model is outlined in

Table 3. Detailed classification report for fingerprint recognition.

Class	Precision	Recall	F1-Score	Support
Real	0.99%	100%	0.99%	2814
Fake	100%	0.99%	0.99%	2895
Accuracy	99.42%			5709
Macro avg	0.99%	0.99%	0.99%	5709
Weighted avg	0.99%	0.99%	0.99%	5709

, providing insights into precision, recall, and F1-score.

Table 4. Comparison of execution time and computational complexity in biometric authentication models.

No	Title	Method	Recognition Accuracy	Execution Time (s)
1	Radzi et al. (2024) [10]	LeNet-5 model	95.8%	1.2
2	Das et al. (2023) [11]	CNN (5 convolutional layers)	96%	1.8
3	Fairuz et al. (2023) [12]	Transfer learning (AlexNet)	95.2%	2.1
4	Su et al. (2024) [13]	LSTM + center ranked loss	97%	5.3
5	Mohammed et al. (2024) [14]	LSTM	94.5%	4.8
6	Yang et al. (2023) [15]	CNN + LSTM	93.5%	3.9
7	Jang et al. (2024) [16]	CNN-LSTM	93.8%	3.6
8	Minaee et al. (2023) [17]	ResNet50 + LSTM	95.7%	4.5
9	Wang and Yan (2024) [24]	CNN-LSTM ensemble (bagging)	88.2%	6.2
10	Proposed method	Hybrid model (CRNN)	99.17%	1.85

provides a detailed illustration of the comparison in execution time and computational complexity across various biometric authentication models, including the proposed model. Additionally, Figure 5, Figure 6, Figure 7 and Figure 8. illustrate the trends of training and validation accuracy, training and validation loss, as well as the precision, recall, and Receiver Operating Characteristic (ROC) curve, highlighting the performance of the model during both training and testing phases.

The proposed CRNN model exhibited robust performance across various datasets, demonstrating its effectiveness in fingerprint classification.

For the Config dataset, the model achieved a training accuracy of 99.42% and a validation accuracy of 99.23%, with corresponding loss values of 0.0124 and 0.0276, respectively. The model also demonstrated exceptional precision, recall, and F1-scores, each exceeding 99%, indicating a high level of accuracy in distinguishing between genuine and fake fingerprints. The confusion matrix revealed minimal misclassifications, further confirming the model’s reliability and effectiveness in biometric authentication tasks.

Meanwhile, for the FVC2000 dataset, the model attained a training accuracy of 99.0% and a validation accuracy of 98.85%, with loss values of 0.035 and 0.040, respectively. The precision, recall, and F1-scores for both real and fake fingerprint classes were consistently above 98%, showcasing the model’s strong ability to classify fingerprints accurately. The confusion matrix revealed a low rate of misclassifications, suggesting that the model is highly effective in correctly identifying both genuine and altered fingerprints.

For the L3-SF-V2 dataset, the model achieved a training accuracy of 98.22% and a validation accuracy of 98.05%, with corresponding loss values of 0.050 and 0.055, respectively. Precision, recall, and F1-scores exceeded 97% for both real and fake fingerprint classes, highlighting the model’s consistent performance across diverse datasets. The confusion matrix further corroborated the model’s effectiveness, with minimal misclassifications, ensuring reliable performance in fingerprint verification tasks.

Table 2 shows a comparison between the proposed hybrid CRNN model and the existing fingerprint classification approaches, demonstrating its superior performance with 99.17% accuracy. Traditional CNN-based models, such as LeNet-5 (95.8%) and a CNN with five convolutional layers (96%), perform well but lack sequential feature learning. Transfer learning models, like AlexNet (95.2%) and ResNet50 + LSTM (95.7%), benefit from pre-trained features but may not fully adapt to fingerprint-specific patterns. Standalone LSTM models (94.5%) struggle with spatial feature extraction, while CNN-LSTM models from [15] (93.5%) and [16] (93.8%) improve sequence learning but still fall short in accuracy. Su et al.’s model (97%) achieves high accuracy but may risk overfitting. The CNN-LSTM ensemble (88.2%) shows that excessive model complexity can reduce performance.

Dataset Bias and Generalization

While the model demonstrates strong performance, as indicated by the quantitative metrics, a deeper analysis of the confusion matrix reveals certain misclassifications, particularly with altered fingerprints, such as those affected by ridge thinning or obliteration. These alterations obscure critical minutiae points, causing the model to misclassify them as either genuine fingerprints or other types of altered prints. However, it is essential to acknowledge that these misclassifications are more reflective of the limitations inherent in the dataset, rather than the architecture of the model itself.

The SOCOFing dataset, although comprehensive in terms of sample size and synthetic alterations, primarily uses artificial modifications that may not accurately simulate real-world spoofing techniques or the natural degradation of fingerprints over time. Additionally, the dataset’s focus on fingerprints from individuals of African descent may limit the diversity of the samples and affect the model’s ability to generalize across broader populations. Therefore, the observed misclassifications should be viewed as challenges stemming from the dataset’s nature rather than a failure of the model.

Despite these challenges, the model’s ability to distinguish between real and fake fingerprints remains highly reliable, as demonstrated by the high precision and recall values shown in Table 3. To further validate the significance of the observed performance improvements, a paired t-test across 10 independent runs was performed, yielding a p-value below 0.05, confirming that the accuracy improvements are statistically significant.

In light of these factors, it is clear that while the model performs well with the current dataset, there is potential for even greater performance with more diverse and real-world data. Future work that incorporates a wider range of fingerprint variations and distortions will likely improve the model’s generalization, making it more applicable to practical biometric systems across various demographic groups and real-world conditions.

Table 3 presents the detailed classification performance of the proposed CRNN model, evaluating its effectiveness in distinguishing between real and fake fingerprints. The results indicate a highly balanced performance across both classes, with precision, recall, and F1-score values all exceeding 99%. The real fingerprint class achieves a precision of 99% and recall of 100%, meaning that all real fingerprints were correctly identified, with no false negatives. Similarly, the fake fingerprint class attains 100% precision and 99% recall, ensuring minimal misclassification.

The overall accuracy of 99.42% confirms the model’s robust generalization ability, effectively reducing misclassification errors. The macro and weighted averages across all metrics remain consistent at 99%, further demonstrating the model’s stability in fingerprint classification.

4.5. Performance Analysis

The CRNN model demonstrates high accuracy and robustness in fingerprint classification, as reflected in its balanced precision, recall, and F1-score. The training and validation curves show stable learning with minimal overfitting, while the loss trends confirm efficient convergence. Compared to existing methods, the hybrid CNN-LSTM approach enhances spatial and sequential feature extraction, leading to superior classification performance. These results validate the model’s effectiveness for biometric authentication.

The performance results shown in Table 1 demonstrate the effectiveness of the proposed hybrid CNN-LSTM model in fingerprint classification on the original CONFIG dataset. The negligible difference between training (99.42%) and validation (99.23%) accuracies, along with the low loss gap (0.0124 vs. 0.0276), suggests strong generalization capabilities and minimal overfitting. Similarly, for the FVC2000 dataset, the results shown in Table 1 indicate that the proposed model achieves a training accuracy of 99.0% and a validation accuracy of 98.85%. The minimal difference between training and validation performance, alongside the low loss values (0.035 vs. 0.040), further confirms the model’s robustness and its ability to generalize well without significant overfitting. In the case of the L3-SF-V2 dataset, the model achieves a training accuracy of 98.22% and a validation accuracy of 98.05%. Despite a slightly larger gap compared to the other datasets, the results remain consistent, with low loss values (0.050 vs. 0.055), indicating that the model maintains strong generalization capabilities even across different fingerprint data distributions.

The use of five convolutional layers allowed the model to progressively learn from low-level features such as edges and textures to high-level abstractions like ridge flow and minutiae structures. This deep spatial feature extraction was complemented by LSTM layers, which successfully captured sequential dependencies and contextual information across the fingerprint’s structure. This dual capability enabled the model to effectively handle fingerprint variations, partial prints, and distortions often encountered in real-world biometric systems.

Compared to traditional CNN-only or LSTM-only models, the proposed hybrid approach demonstrated superior adaptability and accuracy. A paired t-test performed over 10 independent runs confirmed that the improvement in classification accuracy was statistically significant (p < 0.05), reinforcing the reliability of the results.

Additionally, the model’s balanced performance across all metrics—accuracy, precision, recall, and F1-score—indicates its ability to classify both real and fake fingerprints without bias, a common issue in many biometric models. These findings validate the proposed model’s robustness, making it suitable for high-security applications such as law enforcement, border control, and financial access systems.

This performance trend is further illustrated in Figure 5, which shows the accuracy and loss curves for both training and validation phases, confirming the model’s stability throughout the learning process.

The performance comparison in Table 2 highlights the superior accuracy of the proposed CNN-LSTM model over existing fingerprint classification approaches. This performance advantage can be attributed to several architectural and methodological improvements.

While previous models such as those by Yang et al. [15] and Jang et al. [16] employed shallow configurations with limited layers, our model incorporates five convolutional layers for deeper spatial feature extraction and three LSTM layers for enhanced sequential modeling. This deeper architecture enables the capture of more complex fingerprint patterns, particularly in the presence of noise or distortion.

In addition, the proposed model benefits from comprehensive data augmentation techniques—including rotation, flipping, and zooming—which improve generalization and robustness to spoofed inputs, an aspect that was not extensively addressed in the cited studies. Unlike Su et al. [13], whose model achieved 97% accuracy but showed potential overfitting due to limited regularization, our approach incorporates dropout, L2 regularization, and early stopping, leading to improved generalization and training stability.

Furthermore, the use of the SOCOFing dataset with synthetic alterations across varying difficulty levels provides a more realistic evaluation setting than those used in previous works, further validating the model’s applicability in real-world biometric authentication systems.

By effectively combining CNNs for spatial feature learning and LSTMs for sequential modeling, the proposed model demonstrates enhanced adaptability, accuracy, and robustness—especially when handling fingerprint distortions and variations—thus confirming its potential for deployment in secure biometric authentication scenarios.

These observations are further supported by Figure 6, which illustrates the precision–recall (PR) curves for each class, indicating strong discrimination ability across all fingerprint categories.

Unlike traditional CNN-only models, which may struggle with distorted or partial fingerprints, the CRNN approach effectively captures both spatial and sequential dependencies, leading to improved classification reliability. This hybrid architecture ensures that the model is better equipped to handle the intricate variations present in real-world fingerprint data, including distortions and partial prints.

The high recall and precision values, as presented in Table 3, suggest that the model does not exhibit bias towards either class, which is a common challenge in fingerprint recognition due to variations in print quality and acquisition conditions. The model’s ability to correctly identify both real and fake fingerprints with minimal misclassification further enhances its credibility as a reliable biometric authentication tool. The execution time and computational complexity results, as illustrated in Table 4, highlight the efficiency of the proposed model compared to other biometric authentication approaches. Despite achieving a significantly higher recognition accuracy, the proposed method maintains a competitive execution time of 1.85 seconds. This balance between speed and accuracy underscores the model’s practical applicability in real-time fingerprint recognition scenarios, where rapid and reliable decision-making is essential.

These results validate the model’s ability to extract meaningful fingerprint features while maintaining high accuracy, robustness, and efficiency, confirming its suitability for practical biometric systems in real-world applications.

The model’s classification performance is further illustrated in Figure 7, where the confusion matrix confirms the model’s ability to accurately distinguish between real and fake fingerprint classes with negligible misclassification.

The confusion matrix in Figure 7 illustrates the classification performance of the proposed CNN-LSTM model for fingerprint authentication. The diagonal values represent correctly classified samples, where 2802 real fingerprints and 2874 fake fingerprints were accurately identified. The off-diagonal values indicate misclassifications, with 2 real fingerprints misclassified as fake and 12 fake fingerprints misclassified as real. The extremely low misclassification rate highlights the model’s strong generalization ability and its effectiveness in distinguishing between real and fake fingerprints. The high precision and recall scores indicate that the model minimizes false positives and false negatives, crucial for ensuring biometric security. The combination of the CNN for deep feature extraction and LSTM for sequential learning contributes to this high classification accuracy, reducing errors in challenging fingerprint patterns.

4.6. Illustrative Results of the Proposed Technique

Figure 8 presents a sample output generated by the proposed technique, effectively showcasing the model’s ability to classify “Real” and “Fake” classes with remarkable precision and reliability. This visualization highlights the robustness of the model in achieving perfect performance, as evidenced by the metrics discussed. By demonstrating the model’s capacity to accurately differentiate between the two classes, serves as a clear and compelling representation of the technique’s success in addressing the classification task with exceptional effectiveness.

4.7. Execution Time Efficiency of the Proposed Model for Real-Time Prediction

As shown in Table 4, the execution time of 1.85 s reflects the efficiency and performance of the proposed model, demonstrating its ability to make predictions swiftly and accurately compared to existing methods,. This quick response time is an essential factor in evaluating the model’s effectiveness. It indicates that the model has been trained efficiently, with optimized algorithms and computational techniques that significantly enhance the speed of predictions.

The actual execution time was calculated using the following formula:

$${\mathbf{T}}_{\mathrm{execution}}={\mathbf{T}}_{\text{prediction\_end}}-{\mathbf{T}}_{\text{input\_start}}$$

(1)

where:

• ${\mathbf{T}}_{\text{input\_start}}$ is the time when the input data (image) is fed into the model.
• ${\mathbf{T}}_{\text{prediction\_end}}$ is the time when the model returns the prediction result (whether real or fake).

The formula captures the complete prediction pipeline duration from input to output. This 1.85 s execution time demonstrates the model’s optimal balance between speed and accuracy, crucial for real-time applications like fingerprint authentication. The efficient processing enables reliable deployment in production environments requiring both rapid response (e.g., border control systems) and high throughput for large datasets.

5. Conclusions and Future Scope

This study proposed a novel Convolutional Recurrent Neural Network (CRNN) model designed to enhance fingerprint classification accuracy and efficiency, distinguishing real from altered fingerprint images. The model integrates Convolutional Neural Networks (CNNs) for feature extraction with Long Short-Term Memory (LSTM) layers to capture temporal dependencies, improving generalization to unseen data. Evaluating the model on the SOCOFing dataset demonstrated high classification accuracy, surpassing traditional methods and proving its suitability for security and biometric authentication applications. Additionally, the model’s execution time was optimized, enabling real-time predictions, making it ideal for applications requiring fast decisions, such as fingerprint verification systems. Despite these successful results, there is room for future improvement. Future research could focus on enhancing the model’s robustness by incorporating diverse datasets to better capture a variety of fingerprint variations. Moreover, exploring techniques like transfer learning and attention mechanisms could improve performance further. Further optimization of the network architecture, along with efforts to reduce computational load, would facilitate real-time processing in low-resource environments. This work establishes a foundation for future research in biometric authentication systems, with potential extensions to other biometric modalities, such as face recognition, iris scanning, or voice identification. These advancements would contribute to the development of multi-modal biometric systems in security.