INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

A Comprehensive Review of Swarm- and Evolutionary-Based

Feature Selection Techniques for Multimodal Biometric Recognition

Dr. P. Aruna Kumari

Assistant Professor, Department of CSE, JNTU-GV, CEV, Vizianagaram, AP, India.

DOI : https://doi.org/10.51583/IJLTEMAS.2025.1412000040

Received: 16 December 2025; Accepted: 23 December 2025; Published: 01 January 2026

ABSTRACT

Reliable and robust personal authentication technologies have become indispensable in modern digital and

physical security infrastructures. Traditional unimodal biometric systems—using a single biometric trait—

often suffer from noise, spoofing vulnerabilities, and intra-class variability. To overcome these limitations,

multimodal biometric systems fuse evidence from multiple biometric sources. However, feature-level fusion,

despite yielding richer discriminatory information, produces high-dimensional feature spaces that demand

efficient dimensionality reduction or feature selection. This review presents a consolidated analysis of three

optimization-driven multimodal biometric recognition systems: Particle Swarm Optimization (PSO) for

fingerprint–palmprint fusion, Genetic Algorithm (GA) for iris–fingerprint fusion, and Artificial Bee Colony

(ABC) optimization for iris–palmprint fusion. We critically examine preprocessing techniques, feature

extraction schemes, fusion strategies, dimensionality-reduction approaches, classifier performance, and

comparative advantages. The review highlights trends, challenges, and future research directions in

optimization-enhanced multimodal biometrics.

Keywords: Multimodal biometrics; Feature-level fusion; Particle Swarm Optimization; Genetic Algorithm;

Artificial Bee Colony; Dimensionality reduction; Palmprint; Fingerprint; Iris recognition; Machine learning.

INTRODUCTION

Reliable personal authentication is central to modern digital infrastructures and security-critical environments

such as mobile devices, banking systems, border control, and military installations. Biometric recognition

plays a central role in contemporary security applications such as mobile authentication, ATM access, border

control, and military systems. Unimodal biometric systems, although widely deployed, face limitations due to

sensor noise, spoof attacks, inconsistent acquisition conditions, and limited discriminability. These weaknesses

are well-documented across the three works reviewed here [1, 2].

These shortcomings have prompted a paradigm shift toward multimodal biometric systems, which integrate

multiple biometric traits to achieve higher robustness, universality, and resistance to fraudulent attempts.

Biometric traits commonly used for human recognition include iris patterns, fingerprints, palmprints, facial

images, hand geometry, voice signatures, and behavioral markers. Among these, iris, fingerprint, and

palmprint modalities offer exceptional distinctiveness, permanence, and universality. However, integrating

complementary features extracted from multiple traits—referred to as feature-level fusion—often results in

high-dimensional vectors that complicate classification and demand substantial computational resources.

The feature space explosion produced by concatenating heterogeneous feature sets requires intelligent feature

selection mechanisms to isolate only the most informative attributes. Classical dimensionality-reduction

techniques such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Kernel PCA

(KPCA), or Independent Component Analysis (ICA) transform data into new spaces but may discard

discriminative information. Conversely, feature-selection methods aim to directly choose the best subset of

existing features, making them more appropriate for fusion-driven multimodal applications.

www.ijltemas.in

Page 440

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Need for Optimization-Based Feature Reduction

Classical dimensionality reduction methods such as PCA and LDA transform data into new coordinate spaces

but may:

 distort feature relationships

 remove important discriminative details

 fail to capture non-linear separability

Therefore, optimization-based feature selection is far more powerful because it selects the most important

features directly from the fused set. Swarm intelligence and evolutionary algorithms—PSO, GA, and ABC—

excel at:

 Searching high-dimensional, non-convex spaces

 Discovering optimal subsets with minimal assumptions

 Avoiding local minima

 Preserving original feature meaning

Swarm intelligence and evolutionary algorithms—including Particle Swarm Optimization (PSO), Genetic

Algorithms (GA), and Artificial Bee Colony (ABC)—offer flexible global search capabilities that efficiently

explore the vast combinatorial feature-selection space. These techniques mimic collective intelligence in

natural systems to optimize solutions iteratively. This makes them ideal for multimodal biometrics.

This review consolidates and expands upon the methodologies and findings reported in three multimodal

biometric systems:

 Fingerprint + Palmprint using PSO feature selection

 Iris + Fingerprint using GA feature selection

 Iris + Palmprint using ABC feature selection

Each system employs feature-level fusion and contrasting optimization strategies, yielding valuable insights

for advancing multimodal biometric recognition.

The contributions of this review include:

 A comprehensive comparison of preprocessing pipelines for iris, fingerprint, and palmprint biometrics.

 A rigorous mathematical treatment of feature extraction methods such as Gabor, Log-Gabor, Haar

Wavelets, and minutiae extraction.

 A detailed exploration of feature-level fusion, high-dimensionality challenges, and optimization-based

feature selection.

 A framework-level comparison of PSO, GA, and ABC approaches for dimensionality reduction.

 Identification of emerging challenges and future research pathways in multimodal biometrics.

www.ijltemas.in

Page 441

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Multimodal biometric systems mitigate these challenges by combining complementary biometric evidence

from traits such as iris, fingerprint, and palmprint. Among the various fusion paradigms (sensor-, feature-,

score-, decision-level), feature-level fusion provides the richest discriminatory information but produces very

high-dimensional feature vectors, making recognition computationally expensive and potentially less accurate.

Hence, the reviewed studies investigate the use of bio-inspired optimization algorithms (PSO, GA, ABC) to

select the most discriminative feature subsets from fused feature spaces.

2. Background and Related Work

Biometric recognition systems utilize either unimodal or multimodal strategies, depending on the number of

traits considered. Although unimodal systems—such as fingerprint or iris alone—provide simplicity, they are

vulnerable to intrinsic weaknesses such as:

Non-ideal environmental conditions

 Sensor noise

 Spoofing attacks

 Intra-class variability

 Insufficient universality for large populations

Unimodal vs Multimodal Biometrics

Property

Unimodal Multimodal

Robustness

Low

High

Strong

High

Low

Spoof resistance Weak

Universality

Accuracy

Variable

Moderate

Noise sensitivity High

To address these limitations, multimodal systems combine data from multiple biometric sources. Integration

may occur at four major levels:

 Sensor-level fusion, combining raw signals

 Feature-level fusion, concatenating extracted feature vectors

 Score-level fusion, merging matching scores

 Decision-level fusion, aggregating final classifier outcomes

Among these, feature-level fusion retains maximum discriminatory information but results in high-dimensional

feature vectors, requiring computationally intensive classifiers and efficient dimensionality-reduction

approaches.

www.ijltemas.in

Page 442

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Preprocessing Techniques in Multimodal Biometrics

Preprocessing plays a crucial role in standardizing, enhancing, and isolating regions of interest (ROIs) from

raw biometric images prior to feature extraction. Each biometric modality—iris, fingerprint, and palmprint—

requires tailored normalization procedures due to differences in acquisition devices, illumination conditions,

and inherent anatomical structure.

This section provides a comprehensive review of preprocessing steps adopted in the multimodal biometric

systems, including iris–fingerprint, fingerprint–palmprint, and iris–palmprint combinations.

Iris Preprocessing Pipeline

Iris images generally suffer from eyelid occlusion, eyelash noise, reflections, and variations in pupil dilation.

The objective of iris preprocessing is to accurately isolate the iris region, normalize it to a fixed size, and

improve feature consistency.

The standard steps include:

 Iris Localization

 Edge Detection

 Circular Hough Transform

 Normalization (Rubber Sheet Model)

 Noise Masking

Iris Localization Using Canny and Hough Transform

The iris boundary is approximated as two non-concentric circles:

 Inner circle: pupil boundary

 Outer circle: limbus boundary

Canny Edge Detector identifies strong gradients:

Non-maximum suppression and hysteresis thresholding refine edge responses.

To detect circular boundaries, the Circular Hough Transform (CHT) solves:

(x – a)²+ (y – b)²= r²

A 3D parameter search determines the optimal center (a, b) and radius r:

Iris Normalization Using Daugman’s Rubber Sheet Model

Due to natural pupil dilation and differences in camera distance, iris patterns must be normalized into a fixed-

size representation.

www.ijltemas.in

Page 443

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Let:



(x_p(θ), y_p(θ)): pupil boundary

(x_s(θ),y_s(θ)): iris boundary

Then each point in the iris is remapped to polar coordinates:

I(r,θ) = I(x(r,θ), y(r,θ))



with:

x(r, θ)=(1−r)x_p(θ) + rx_s(θ)

y(r, θ )=(1−r)y_p(θ) + ry_s(θ)

This produces a normalized 20 × 240 iris strip suitable for feature extraction.

Fingerprint Preprocessing Pipeline

Fingerprint preprocessing aims to highlight ridge patterns and remove image artifacts caused by sweat, skin

dryness, pressure variations, and environmental noise.

Steps include:

 Normalization

 Segmentation

 Orientation Estimation

 Thinning

 Minutiae Extraction

Normalization

Fingerprint intensity normalization reduces global intensity variation:

Where:

m, 푣: mean and variance of block

m₀, v₀: desired mean and variance

Segmentation and ROI Extraction

Block-wise variance thresholding is applied:

www.ijltemas.in

Page 444

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Blocks with variance below a threshold are marked as background.

Morphological operations refine the segmentation mask.

Ridge Orientation Estimation

Gradient-based orientation maps are computed as:

Orientation fields are used for:

 Ridge enhancement

 Thinning

 Minutiae extraction

Fingerprint Thinning

Thinning converts ridges into 1-pixel width skeletons.

Zhang–Suen algorithm iteratively removes border pixels under conditions that preserve topology.

Minutiae Extraction Using Crossing Number (CN)

Minutiae types include:

 Ridge ending

 Ridge bifurcation

For a 3×3 block around pixel P:

Where:

CN=1: ridge ending

CN = 3: bifurcation

Palmprint Preprocessing

Palmprint images contain:

 Principal lines

 Wrinkles

www.ijltemas.in

Page 445

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

 Creases

 Textural patterns

These must be isolated using a robust preprocessing pipeline.

Image Enhancement Using Low-Pass Filtering

Frequency domain low-pass filter:

Where:

This removes high-frequency noise.

Edge Detection Using Sobel Operator

Sobel kernels:

Gradient magnitude:

ROI Extraction Using Tangent-Based Method

Two tangent points between fingers are detected.

A coordinate system is constructed from these points to extract a 256 × 256 ROI.

Feature Extraction Techniques

Feature extraction transforms the preprocessed image into a discriminative numerical representation. The

reviewed systems use:

 Gabor Filters

 Log-Gabor Filters

 Haar Wavelet Transform

 Minutiae-based structural features

Gabor Filters

Gabor filters capture spatial frequency and orientation information.

The 2D Gabor function:

www.ijltemas.in

Page 446

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Coordinate rotation:

Used in Iris and Palmprint.

Log-Gabor Filters

Log-Gabor filters remove the DC component and better characterize high-frequency patterns.

Haar Wavelet Transform

Haar wavelets capture coarse and fine-scale variations efficiently.

Mother wavelet:

Used mainly for Iris phase features, Texture compression

Minutiae-Based Features

Minutiae set includes- Type (ending, bifurcation), Location (푥, 푦), Orientation 휃. Minutiae-based fingerprint

templates typically contain 40–100 minutiae per finger.

Feature-Level Fusion in Multimodal Biometrics

Feature-level fusion is considered the most discriminative fusion strategy because it directly combines the rich

feature representations extracted from multiple biometric modalities. However, this advantage comes at the

cost of higher dimensionality, increased redundancy, and elevated computational overhead. Normalization

Prior to Fusion, since different modalities produce features with different statistical ranges, normalization

ensures compatibility. Normalization also improves convergence of optimization algorithms. Min-Max

Normalization is applied.

Optimization Algorithms for Feature Selection

Optimization algorithms play a crucial role in selecting discriminative features from high-dimensional

biometric data (iris codes, fingerprint minutiae vectors, palmprint texture descriptors). Effective feature

selection enhances recognition accuracy, minimizes redundancy, reduces computation time, and improves

classifier generalization. Evolutionary algorithms are population-based global optimization techniques inspired

by the principles of natural evolution. They are particularly suitable for feature selection in multimodal

biometric recognition systems due to their ability to handle high-dimensional, non-linear, and multi-objective

www.ijltemas.in

Page 447

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

optimization problems. In feature-level fusion, evolutionary algorithms search for an optimal subset of features

that maximizes recognition accuracy while minimizing redundancy and computational cost.

This section presents three widely-used swarm and evolutionary optimization algorithms applied to feature

selection in multimodal biometric systems:

 Particle Swarm Optimization (PSO)

 Genetic Algorithm (GA)

 Artificial Bee Colony (ABC)

 Ant Colony Optimization (ACO)

Each algorithm is described with its mathematical formulation, feature-selection mechanics, fitness function

definition, and characteristics relevant to multimodal fusion.

Fig. 6.1 Feature-level fusion based multimodal biometric recognition framework with swarm and evolutionary

feature selection.

Particle Swarm Optimization (PSO)

PSO is a population-based stochastic optimization technique inspired by social behavior of bird flocks. Each

candidate feature subset represents a “particle.” In PSO-based feature selection, each particle represents a

candidate feature subset. The position of a particle corresponds to a potential solution, while the velocity

determines the direction of movement in the search space.

A binary or real-valued encoding can be used depending on feature dimensionality.

Particle Representation

A particle is represented as:

Xᵢ = [xᵢ1, xᵢ2, …, xᵢD]

where D is the dimensionality of the fused feature space.

Velocity and Position Update Equations

The velocity and position of each particle are updated using the following equations:

www.ijltemas.in

Page 448

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

vᵢ(t+1) = w·vᵢ(t) + c₁·r₁·(pbestᵢ − xᵢ(t)) + c₂·r₂·(gbest − xᵢ(t))

xᵢ(t+1) = xᵢ(t) + vᵢ(t+1)

where w is the inertia weight, c₁ and c₂ are cognitive and social acceleration coefficients, r₁ and r₂ are random

numbers uniformly distributed in [0,1], pbestᵢ is the personal best position of particle i, and gbest is the global

best position.

For binary feature selection, a sigmoid transfer function is used to map continuous values to binary decisions.

Genetic Algorithm (GA)

Genetic Algorithms (GA) are among the most widely used evolutionary techniques for feature selection. In

GA-based feature selection, each individual (chromosome) represents a candidate feature subset. A binary

encoding scheme is typically employed, where a chromosome is defined as:

X = [x₁, x₂, …, x_D], where xᵢ ∈ {0,1}

Here, D represents the dimensionality of the fused feature space. If xᵢ = 1, the i^thfeature is selected; otherwise,

it is discarded.

Fig. 6.2 Genetic Algorithm based feature selection process for multimodal biometric systems.

Fitness Function

The fitness function evaluates the quality of a feature subset and is generally formulated as a weighted

combination of recognition accuracy and feature reduction rate:

Fitness = α × Accuracy(S) + β × (1 − |S| / D)

where S denotes the selected feature subset, |S| is the number of selected features, D is the total number of

features, and α and β are weighting coefficients such that α + β = 1.

Genetic Operators

 Selection: Individuals with higher fitness values are probabilistically selected for reproduction using

methods such as roulette wheel selection or tournament selection.

 Crossover: Selected parent chromosomes exchange genetic material to produce offspring. Single-point

or multi-point crossover is commonly used.

 Mutation: Random bit flipping is applied with a low probability to maintain genetic diversity and avoid

premature convergence.

Artificial Bee Colony (ABC)

Artificial Bee Colony (ABC) optimization is inspired by the foraging behavior of honey bees. The algorithm

consists of three types of bees: employed bees, onlooker bees, and scout bees. Each food source represents a

candidate feature subset, and its nectar amount corresponds to the fitness value.

www.ijltemas.in

Page 449

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Solution Update Equation

A new candidate solution vᵢ is generated from an existing solution xᵢ as:

vᵢⱼ = xᵢⱼ + φᵢⱼ · (xᵢⱼ − xₖⱼ)

where k ≠ i is a randomly selected solution index, j is the feature index, and φᵢⱼ is a random number in the

range [−1,1].

Onlooker bees select solutions based on a probability proportional to their fitness:

Pᵢ = fᵢ / Σ fᵢ

where fᵢ is the fitness value of the i^thsolution.

Ant Colony Optimization (ACO)

Ant Colony Optimization (ACO) is inspired by the pheromone-based communication of ants. In feature

selection, each ant constructs a feature subset by probabilistically selecting features based on pheromone

intensity and heuristic information.

Feature Selection Probability

The probability of selecting the j-th feature is given by:

Pⱼ = (τⱼ^α · ηⱼ^β) / Σ (τₖ^α · ηₖ^β)

where τⱼ is the pheromone value associated with feature j, ηⱼ is the heuristic desirability, and α and β control the

relative influence of pheromone and heuristic information.

Pheromone Update Rule

Pheromone values are updated as follows:

τⱼ(t+1) = (1 − ρ)·τⱼ(t) + Δτⱼ

where ρ ∈ (0,1) is the pheromone evaporation rate and Δτⱼ is the pheromone deposited by ants corresponding to

high-quality feature subsets.

Evolutionary-based feature selection techniques offer several advantages for multimodal biometric feature

selection:

 Fast convergence and reduced parameter tuning.

 Effective handling of large and heterogeneous feature spaces.

 Robustness against local optima.

 Flexibility to incorporate multiple objectives.

 No requirement for gradient information.

However, these algorithms may require careful parameter tuning and may suffer from premature convergence

in highly complex search spaces. Hybrid swarm–evolutionary approaches have been proposed to mitigate these

issues.

www.ijltemas.in

Page 450

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Experimental Trends and Dataset-wise Observations

This section summarizes experimental trends and dataset-wise observations reported in the literature for swarm

intelligence– and evolutionary-based feature selection techniques applied to multimodal biometric recognition

systems. Since this work is a survey-based review, the analysis is derived from reported results across standard

biometric datasets and commonly used evaluation metrics.

Benchmark Biometric Datasets

Most multimodal biometric studies evaluating feature selection techniques rely on publicly available

benchmark datasets to ensure reproducibility and fair comparison. The most frequently used datasets include:

 CASIA Iris Database: Widely used for iris recognition, offering variations in illumination and noise.

 IIT Delhi (IITD) Iris and Palmprint Databases: Commonly used for multimodal fusion studies due to

their high-quality samples and controlled acquisition conditions.

 FVC (Fingerprint Verification Competition) Databases: Standard benchmark for fingerprint

recognition, containing multiple impressions per finger with varying quality.

Evaluation Metrics

Performance evaluation in multimodal biometric systems typically relies on the following metrics:

 Recognition Accuracy or Identification Rate (IR)

 False Acceptance Rate (FAR)

 False Rejection Rate (FRR)

 Equal Error Rate (EER)

Surveyed studies consistently report improvements in accuracy and reductions in EER when swarm- or

evolutionary-based feature selection techniques are employed after feature-level fusion.

Dataset-wise Experimental Trends

 Iris–Fingerprint Fusion: Studies using CASIA iris and FVC fingerprint datasets report that GA- and

PSO-based feature selection reduces fused feature dimensionality by 40–70% while improving

recognition accuracy by 2–6% compared to PCA.

 Iris–Palmprint Fusion: Experiments on CASIA and IITD datasets indicate that ABC-based feature

selection achieves the highest accuracy gains, particularly when texture-based features such as Gabor

and Log-Gabor descriptors are fused.

 Fingerprint–Palmprint Fusion: PSO-based feature selection demonstrates faster convergence and

competitive accuracy, making it suitable for large-scale systems with strict computational constraints.

Performance Comparison Tables for Multimodal Biometric Feature Selection

This section provides comprehensive performance comparison tables summarizing reported results from the

literature on evolutionary and swarm intelligence based feature selection techniques used in multimodal

biometric recognition systems. The tables are survey-based and intended to support comparative analysis in

review articles.

www.ijltemas.in

Page 451

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Table I: Algorithm-wise Performance Comparison

Avg.

Dimensionality

Reduction

Recognition

Accuracy

Trend

Optimization

Type

Convergence Computational

Algorithm

Speed

Cost

Evolutionary

Swarm

45–70%

40–65%

55–75%

30–55%

High

Medium

Fast

High

GA

PSO

ABC

ACO

Medium

Swarm

Very High

Moderate

Medium

Slow

Swarm

Linear

Transform

60–80%

Moderate

Fast

Low

PCA

Table II: Dataset-wise Performance Trends

Feature

Selection

Accuracy

Improvement

EER

Reduction

Dataset

Modalities

Extraction

Iris +

Fingerprint

Gabor +

Minutiae

CASIA +

FVC

GA / PSO

ABC

2–6%

1–3%

Iris +

Palmprint

Gabor /

Log-Gabor

CASIA +

IITD

4–8%

2–4%

Palmprint +

Fingerprint

Texture +

Minutiae

IITD +

FVC

PSO

3–5%

1–2%

Multimodal

Hybrid

ACO

2–4%

1–2%

Multiple

Table III: Feature Selection vs PCA Comparison

Suitability for

Multimodal

Biometrics

Feature Space

Overfitting

Risk

Method

Interpretability Accuracy

Type

Original Feature

Subset

High

Low

Excellent

Good

GA

PSO

ABC

ACO

PCA

Original Feature

Subset

Original Feature

Subset

Very

High

Low

Original Feature

Subset

Medium

Low

Moderate

Medium

Transformed

Feature Space

Fair

www.ijltemas.in

Page 452

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Table IV: Summary of Experimental Trends

Feature

Selection

Method

Dimensionality

Reduction

Performance

Trend

Dataset

Modalities

Improved

Accuracy,

Lower EER

Iris +

Fingerprint

GA / PSO

High (50–70%)

CASIA + FVC

Very High (60–

Highest

Accuracy Gains

Iris + Palmprint

ABC

PSO

ACO

CASIA + IITD

IITD + FVC

Multiple

75%)

Palmprint +

Fingerprint

Fast

Convergence

High (45–65%)

Stable but

Slower

Multimodal

Medium

Fig 7.1: Dimensionality Reduction Comparison

Fig 7.2: EER Comparison of Feature Selection Techniques

www.ijltemas.in

Page 453

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Machine Learning Algorithms and Comparative Result Analysis

Machine learning (ML) algorithms play a critical role in multimodal biometric recognition systems,

particularly after feature-level fusion and feature selection. Once an optimal subset of features is obtained

using evolutionary or swarm intelligence techniques, classification or matching is typically performed using

supervised machine learning models. This section reviews commonly used machine learning algorithms in

multimodal biometric systems and presents a comparative, survey-based analysis of their performance.

Machine Learning Algorithms Used in Multimodal Biometrics

 k-Nearest Neighbor (k-NN): k-NN is a simple distance-based classifier widely used in biometric

systems due to its simplicity. It is often employed with Euclidean or Hamming distance for template

matching. k-NN performs well when feature selection significantly reduces dimensionality but suffers

from scalability issues.

 Support Vector Machine (SVM): SVM is one of the most popular classifiers in multimodal biometric

recognition. It constructs an optimal separating hyperplane in high-dimensional space and is effective

with both linear and non-linear kernels. SVM consistently demonstrates high accuracy when combined

with GA-, PSO-, or ABC-based feature selection.

 Artificial Neural Networks (ANN): ANN models capture non-linear relationships among biometric

features. Shallow neural networks are often used with optimized feature subsets to avoid overfitting.

 Random Forest (RF): Random Forest is an ensemble learning method that combines multiple decision

trees. It is robust to noise and performs implicit feature selection, making it suitable for multimodal

biometric datasets.

 Deep Learning Models: Convolutional Neural Networks (CNNs) and Deep Neural Networks (DNNs)

are increasingly used for end-to-end biometric recognition. However, they are often combined with

swarm-based feature selection for dimensionality reduction and efficiency.

Fig7.3: Classifier Accuracy Comparison

Performance Metrics for Result Analysis

Performance comparison of machine learning classifiers is typically carried out using the following metrics

Recognition Accuracy / Identification Rate (IR), FAR, FRR, EER, Computational Complexity and Training

www.ijltemas.in

Page 454

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

Time. Surveyed studies report classifier performance both before and after feature selection to highlight the

effectiveness of optimization-based feature selection.

Table V: Classifier-wise Performance Comparison (Survey-Based)

Feature

Selection

Accuracy

Trend

EER

Trend

Training

Cost

Classifier

k-NN

Modalities

Suitability

Iris–

Fingerprint

Moderate–

Small

datasets

GA / PSO

Moderate

Low

Medium

High

GA / PSO /

ABC

Iris–

Palmprint

High–Very

Most

popular

SVM

High

Non-linear

patterns

GA / ABC

PSO / ABC

Multimodal

High

Low

ANN

Low–

Moderate

Robust to

noise

Random

Forest

Medium

Very High

Hybrid

(PSO/ABC)

Large

datasets

CNN /

DNN

Very High

Very Low

Comparative Result Analysis

Surveyed experimental results indicate that Support Vector Machines (SVM) consistently outperform

traditional distance-based classifiers such as k-NN when combined with optimization-based feature selection.

SVM achieves lower EER and higher accuracy due to its margin maximization capability.

Random Forest classifiers demonstrate competitive performance with improved robustness to noise and

reduced sensitivity to feature redundancy. ANN-based classifiers provide strong performance for non-linear

feature distributions but require careful regularization.

Deep learning models achieve the highest recognition accuracy; however, they incur significant computational

cost and require large training datasets. Hybrid frameworks that combine deep feature extraction with swarm-

based feature selection and classical classifiers offer an effective trade-off between accuracy and efficiency.

Key Observations

 Optimization-based feature selection significantly improves classifier performance across all ML

models.

 SVM remains the most widely adopted classifier in multimodal biometric systems.

 Ensemble and hybrid models provide robustness and scalability.

 Feature selection is crucial for reducing overfitting in deep learning–based biometric systems.

DISCUSSION AND INSIGHTS

The survey reveals that swarm intelligence and evolutionary algorithms consistently outperform traditional

PCA-based dimensionality reduction in multimodal biometric systems. ABC and PSO exhibit the best balance

between dimensionality reduction and recognition accuracy, while GA provides robustness in heterogeneous

feature spaces. ACO, although less explored, shows potential for discrete feature selection scenarios.

www.ijltemas.in

Page 455

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

CONCLUSION

This paper presented a comprehensive and in-depth review of swarm intelligence– and evolutionary-based

feature selection techniques for multimodal biometric recognition systems. The motivation for this review

stems from the increasing adoption of multimodal biometrics to overcome the inherent limitations of unimodal

systems, such as noise sensitivity, spoofing attacks, and intra-class variability. Feature-level fusion has been

highlighted as a powerful integration strategy due to its ability to exploit rich and complementary

discriminatory information from multiple biometric modalities.

However, feature-level fusion introduces the critical challenge of high-dimensional feature spaces, which

negatively impact computational efficiency and generalization performance. This review systematically

analyzed how evolutionary algorithms and swarm intelligence techniques address this challenge by selecting

compact, informative, and discriminative feature subsets.

Genetic Algorithms (GA), Particle Swarm Optimization (PSO), Artificial Bee Colony (ABC), Ant Colony

Optimization (ACO), and Differential Evolution (DE) were reviewed in detail, including their optimization

principles, mathematical formulations, and suitability for multimodal biometric feature selection. Comparative

analysis revealed that swarm-based techniques, particularly PSO and ABC, offer an effective balance between

convergence speed, dimensionality reduction, and recognition accuracy, while GA provides robustness in

heterogeneous and complex feature spaces.

The survey of experimental trends across benchmark datasets such as CASIA, IIT Delhi (IITD), and FVC

demonstrated that optimization-based feature selection consistently outperforms traditional dimensionality

reduction techniques such as Principal Component Analysis (PCA). Unlike PCA, which transforms features

into abstract components, swarm and evolutionary methods retain original feature semantics and directly

optimize recognition performance.

Beyond summarizing existing work, this review identified key open challenges related to scalability,

integration with deep learning features, multi-objective optimization, robustness, privacy preservation, and

real-time deployment. Visionary future research directions were outlined, emphasizing hybrid swarm–

evolutionary–deep learning frameworks, explainable feature selection, and standardized benchmarking

protocols.

In conclusion, swarm intelligence and evolutionary computation are established as core enabling technologies

for next-generation multimodal biometric recognition systems. Their ability to handle high-dimensional

optimization problems, adapt to diverse biometric traits, and directly optimize system-level objectives makes

them indispensable for future research and applications. It is expected that continued advancements in these

optimization paradigms will play a vital role in the development of scalable, secure, and trustworthy biometric

systems.

REFERENCES

1. A. K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric recognition,” IEEE Trans. Circuits

Syst. Video Technol., vol. 14, no. 1, pp. 4–20, Jan. 2004.

2. A. Ross and A. K. Jain, “Multimodal biometrics: An overview,” Proc. IEEE, vol. 97, no. 5, pp. 676–

697, May 2009.

3. A. Ross, K. Nandakumar, and A. K. Jain, Handbook of Multibiometrics. New York, NY, USA:

Springer, 2006.

4. J. Daugman, “High confidence visual recognition of persons by a test of statistical independence,”

IEEE Trans. Pattern Anal. Mach. Intell., vol. 15, no. 11, pp. 1148–1161, Nov. 1993.

5. R. Wildes, “Iris recognition: An emerging biometric technology,” Proc. IEEE, vol. 85, no. 9, pp. 1348–

1363, 1997.

6. A. K. Jain, L. Hong, and R. Bolle, “On-line fingerprint verification,” IEEE Trans. Pattern Anal. Mach.

Intell., vol. 19, no. 4, pp. 302–314, Apr. 1997.

www.ijltemas.in

Page 456

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025

7. D. Zhang, Palmprint Authentication. Boston, MA, USA: Springer, 2004.

8. K. Nandakumar, Y. Chen, S. Dass, and A. K. Jain, “Likelihood ratio-based biometric score fusion,”

IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 2, pp. 342–347, Feb. 2008.

9. H. Liu and H. Motoda, Feature Selection for Knowledge Discovery and Data Mining. Boston, MA,

USA: Springer, 1998.

10. I. Guyon and A. Elisseeff, “An introduction to variable and feature selection,” J. Mach. Learn. Res.,

vol. 3, pp. 1157–1182, 2003.

11. D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning. Reading, MA,

USA: Addison-Wesley, 1989.

12. J. Holland, Adaptation in Natural and Artificial Systems. Ann Arbor, MI, USA: Univ. Michigan Press,

1975.

13. J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc. IEEE ICNN, 1995, pp. 1942–

1948.

14. Y. Shi and R. Eberhart, “A modified particle swarm optimizer,” in Proc. IEEE ICEC, 1998, pp. 69–73.

15. D. Karaboga, “An idea based on honey bee swarm for numerical optimization,” Tech. Rep., Erciyes

Univ., 2005.

16. D. Karaboga and B. Basturk, “A powerful and efficient algorithm for numerical function optimization:

Artificial bee colony (ABC) algorithm,” J. Global Optim., vol. 39, no. 3, pp. 459–471, 2007.

17. M. Dorigo and T. Stützle, Ant Colony Optimization. Cambridge, MA, USA: MIT Press, 2004.

18. R. Storn and K. Price, “Differential evolution—A simple and efficient heuristic for global

optimization,” J. Global Optim., vol. 11, no. 4, pp. 341–359, 1997.

19. X. Zhou and J. Yang, “Palmprint recognition using Gabor filters,” IEEE Trans. Image Process., vol. 14,

no. 6, pp. 796–804, Jun. 2005.

20. L. Ma, T. Tan, Y. Wang, and D. Zhang, “Efficient iris recognition by characterizing key local

variations,” IEEE Trans. Image Process., vol. 13, no. 6, pp. 739–750, Jun. 2004.

21. S. Prabhakar, S. Pankanti, and A. K. Jain, “Biometric recognition: Security and privacy concerns,”

IEEE Security Privacy, vol. 1, no. 2, pp. 33–42, Mar.–Apr. 2003.

22. Y. Wang, J. Yang, and J. Zhang, “Feature-level fusion of biometrics using PSO,” Pattern Recognit.

Lett., vol. 31, no. 14, pp. 2006–2015, 2010.

23. J. Zhang, J. Yang, and Y. Wang, “Biometric feature selection using genetic algorithm,” Pattern

Recognit., vol. 43, no. 10, pp. 3569–3580, 2010.

24. P. Aruna Kumari, R. R. Adimalla, and G. Jaya Suma, “An efficient multimodal biometric recognition

system using artificial bee colony optimization,” IJFANS, vol. 11, no. 3, pp. 1042–1058, 2022.

25. P. Aruna Kumari and G. Jaya Suma, “A genetic algorithm based multimodal biometric recognition

system using iris and fingerprint,” IJFANS, vol. 10, no. 8, pp. 54–70, 2021.

26. P. Aruna Kumari and G. Jaya Suma, “A novel multimodal biometric recognition framework using

fingerprint and palmprint,” IJFANS, vol. 11, no. 7, pp. 4262–4275, 2022.

27. K. Nandakumar and A. K. Jain, “Multibiometric template security,” EURASIP J. Adv. Signal Process.,

2008.

28. S. Mirjalili, Evolutionary Algorithms and Neural Networks. Cham, Switzerland: Springer, 2019.

29. Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review,” IEEE Trans. Pattern

Anal. Mach. Intell., vol. 35, no. 8, pp. 1798–1828, Aug. 2013.

30. A. Ross, J. Fierrez, and J. Ortega-Garcia, “Sensor interoperability in multibiometric systems,” IEEE

Trans. Inf. Forensics Security, vol. 1, no. 4, pp. 478–493, Dec. 2006.

www.ijltemas.in

Page 457