Federated Learning for Surveillance Systems: A Literature Review and AHP Expert-Based Evaluation

Abstract

This study explores the application of federated learning (FL) in security camera surveillance systems to overcome the structural limitations inherent in traditional centralized artificial intelligence (AI) training approaches, while simultaneously enhancing operational efficiency and data security. Conventional centralized AI models require the transmission of raw surveillance data from individual security camera units to a central server for model training, which poses significant challenges, including network congestion, a heightened risk of personal data leakage, and inadequate adaptation to localized environmental characteristics. These limitations are particularly critical in high-security environments such as military bases and government facilities, where reliability and real-time processing are paramount. In contrast, FL enables decentralized training by retaining data on local devices and sharing only model parameters with a central aggregator, thereby improving privacy preservation, reducing communication overhead, and facilitating adaptive, context-aware learning. This paper does not present a new federated learning algorithm or original experiment. Instead, it synthesizes existing research findings and applies the Analytic Hierarchy Process (AHP) to evaluate and prioritize critical factors for deploying FL in surveillance systems. By combining literature-based evidence with structured expert judgment, this study provides practical guidelines for real-world application. This paper identifies four key performance metrics—detection accuracy, false alarm rate, response time, and network load—and conducts a comparative analysis of FL and centralized AI-based approaches in the recent literature. In addition, the AHP is employed to evaluate expert survey data, quantitatively prioritizing eight critical factors for effective FL implementation. The results highlight detection accuracy and data security as the most significant concerns, indicating that FL presents a promising solution for future smart surveillance infrastructures. This research contributes to the advancement of AI-powered surveillance systems that are both high-performing and resilient under stringent privacy and operational constraints.

1. Introduction

In contemporary society, security camera systems are utilized as core security measures for crime prevention and anomaly detection across diverse spaces including public places, military facilities, border areas, and private buildings. As security threats such as terrorism, intrusion, and abnormal behavior detection become increasingly sophisticated, the role of security camera-based surveillance has become even more critical, with emphasis being placed on constructing intelligent surveillance systems that go beyond simple recording-centered monitoring. Recently, the demand for the automation of such surveillance and perimeter security functions has increased dramatically, and consequently, artificial intelligence (AI) technology capable of analyzing large volumes of video data in real-time has been gaining attention [1,2,3]. Particularly, intelligent surveillance technologies capable of high-level analysis, such as object detection, human abnormal behavior recognition, and facial recognition, have emerged as key elements in national security and social safety network construction [4].

However, the existing centralized AI learning approaches contain numerous technical limitations. First, the structure of transmitting large-scale video data collected from individual security camera equipment to central servers for learning inevitably involves data movement, which increases the possibility of personal information leakage or sensitive information exposure during this process. Particularly in highly sensitive locations such as military facilities, hospitals, and schools, privacy protection issues can act as very serious constraint factors [5]. Second, in the process of transmitting real-time data from numerous security camera equipment, network load and transmission delays occur, which impair the system’s real-time analysis capability and response speed [6]. Third, while data collected from various environments contains regional characteristics and context, centralized learning inevitably generates generalized models that have difficulty reflecting these regional characteristics, which can lead to degraded detection performance in actual operational environments [7].

Federated learning (FL) has recently emerged as an alternative technology to overcome these problems. Federated learning operates by training AI models within local terminals without transmitting data externally from each security camera equipment or local node, then transmitting only the weights of the trained models to a central server for integration. In this approach, data is safely stored within the equipment, and since raw data is not leaked externally, it is very advantageous for personal information protection [8]. Additionally, as each equipment learns models individually, customized learning reflecting regional characteristics is possible, and the effects of communication cost reduction and transmission delay minimization can also be expected. From the security aspect, risks of accidents such as server hacking or data leakage can be reduced, and it can facilitate future transitions to reliable, distributed AI structures [9].

Under this background, this study aims to comprehensively analyze the technical and structural advantages that federated learning-based AI models can provide compared to existing centralized learning approaches in security camera surveillance and perimeter security environments. Rather than merely comparing performance metrics, this research conducts an integrated review of factors that must be considered in the actual deployment and operational aspects of surveillance systems, presenting the potential of federated learning across various dimensions including performance (accuracy), efficiency (resource utilization), security (privacy and information protection), and scalability (system application scope). To achieve this objective, this paper first systematically organizes the theoretical differences and technical architectures between conventional AI learning and federated learning, and examines the practical superiority of federated learning through an analysis of previously published research and actual field application cases.

In particular, to derive the core elements necessary for establishing federated learning systems, it is essential to consider that the environments where surveillance systems are deployed are extremely diverse and complex. Since each installation environment exhibits different characteristics in terms of network infrastructure, hardware performance, data characteristics, and security requirements, it is difficult to sufficiently identify the comprehensive core elements required for federated learning system construction based solely on experimental results from specific environments. Therefore, comprehensively collecting opinions from relevant experts who possess diverse field experience and specialized knowledge to derive core elements represents a more practical and effective approach.

From this perspective, this study conducts an Analytic Hierarchy Process (AHP) analysis targeting expert groups to derive the most critical consideration factors and their priorities when designing federated learning-based security camera systems. The core elements derived through the AHP analysis are expected to serve as important guidelines for effectively applying federated learning in various surveillance security environments, minimizing technical risks that may arise during system design and implementation processes, and providing practical guidelines for successful deployment.

Unlike studies that propose new FL algorithms or optimization techniques, the contribution of this work lies in synthesizing validated experimental evidence from the prior literature and translating it into a structured framework for real-world surveillance system design. By integrating these findings with an AHP-based expert evaluation, this study bridges the gap between academic research outcomes and practical decision-making in security-sensitive environments.

The primary contribution of this paper is the application of the AHP to systematically prioritize key factors for federated learning implementation in high-security surveillance environments. While the literature review consolidates findings from prior studies, it mainly serves as a contextual basis to inform and support the expert-driven AHP evaluation. This distinguishes our work from survey-oriented studies, positioning the AHP results as the central focus of the paper.

It is important to note that this study is not intended to propose or experimentally validate a new FL algorithm. Rather, it is designed to integrate validated results from the existing literature with expert-based AHP analysis, thereby producing systematic design guidelines for FL implementation in high-security surveillance environments.

2. Theoretical Background

This chapter examines the transition from traditional security camera systems to intelligent surveillance systems. It analyzes the structural characteristics of existing centralized AI learning approaches and their limitations in large-scale video data transmission, security, and real-time processing, and reviews the concepts and technical structures of federated learning technology to address these issues. Additionally, it explores the technical challenges that arise when implementing federated learning, such as data imbalance, computational performance, communication synchronization, and security threats, and comprehensively investigates the applicability of federated learning in security camera surveillance environments and the expected effects in terms of performance, security, and communication efficiency.

2.1. Development of Intelligent Security Camera Surveillance System

Traditional security camera systems were limited to simply storing video information and requiring direct human monitoring, but with the rapid advancement of artificial intelligence technology, the transition to intelligent surveillance systems is occurring rapidly. This transition means evolving beyond simple recording functions to systems equipped with the capability to analyze large-scale video collected in real-time, automatically identify dangerous situations, and provide warnings. Recently, video analysis technologies integrating various functions such as object recognition, human behavior analysis, facial recognition, license plate tracking, and abnormal behavior detection have been applied, and these systems are playing key roles in areas such as crime prevention, military base security, urban safety, and access control. In particular, these systems are developing from passive surveillance to active and situation-responsive surveillance, with emphasis being placed on the function of recognizing risk factors in real-time and alerting administrators.

As security cameras are required to have the capability to collect video in real-time and immediately analyze it to respond to situations, sophisticated analysis algorithms and powerful learning models are essential. However, since there are many technical and structural constraints in implementing such functions through the existing methods, new approaches to AI learning are required. Since federated learning enables dynamic object detection in surveillance videos, it is possible to enhance real-time threat response and privacy protection by learning models on edge devices without centralizing data [10]. Furthermore, using federated dynamic graph neural networks in distributed surveillance systems enables effective object tracking and privacy protection by aggregating structural and dynamic information without storing data centrally [11].

2.2. Structure and Limitations of Existing AI Learning Method

The existing centralized artificial intelligence learning method involves collecting data from multiple devices, transmitting it to a central server, and then training models based on that data. This approach has the advantage of being able to reflect the statistical characteristics of the entire dataset by learning models based on integrated data.

However, in large-scale distributed environments such as security cameras, several limitations arise. First, video data generated from security cameras has the characteristics of high resolution and large capacity, so continuous data transmission excessively occupies network bandwidth, which causes communication bottlenecks throughout the system and makes real-time analysis difficult. Second, the process of transmitting data collected from sensitive locations such as military facilities or hospitals to external servers can pose security threats and has a high possibility of violating related laws such as personal information protection laws. Third, because precise modeling considering various situations and conditions in physically distributed environments is difficult, a single model learned on a central server may have reduced detection accuracy for specific regions or situations.

These problems become causes that reduce the effectiveness of security camera surveillance systems and emphasize the necessity of more flexible and secure learning methods. Furthermore, the existing centralized learning methods face serious problems, such as privacy protection issues and high data transmission costs, and these problems can be mitigated through the distributed approach of federated learning [4]. Centralized systems often suffer from data heterogeneity and communication overhead, and federated learning can address these issues by supporting local model learning and aggregation [12].

While centralized AI learning provides a useful performance benchmark, it faces fundamental challenges in security-sensitive environments. First, aggregating sensitive surveillance data into a central repository increases the risks of data leakage. Second, centralized training incurs substantial communication overhead when transferring large datasets across constrained or isolated networks. Third, its ability to handle heterogeneous (Non-IID) data sources is limited compared to FL. These limitations reduce the feasibility of centralized approaches in practical deployment scenarios.

2.3. Concept and Technical Structure of Federated Learning

Federated learning is a distributed machine learning approach where model training is performed autonomously on the devices where data is generated, and only the weights or parameters of the trained models are transmitted to the server for integration. Federated learning operates in a manner that ensures data privacy and complies with regulations such as the GDPR (General Data Protection Regulation) by training models locally on devices and aggregating updates to form a global model [13]. This consists of a structure where the data itself remains on local devices, while the central server collects parameters to update the global model. This approach enables collaborative learning without data movement and provides significant advantages especially in environments where privacy protection is important [14]. For example, federated learning dramatically improves privacy protection and security by maintaining raw data within devices without transmitting it externally, and can reduce waste of communication resources. Additionally, since learning occurs on each local device, customized learning reflecting regional characteristics is possible, which consequently leads to improved surveillance accuracy.

The main techniques of federated learning include model compression, asynchronous updates, and robust aggregation, which enhance the scalability and efficiency of various applications [13]. Federated learning continuously improves the global model through the process of periodically integrating models at the server, and algorithms such as FedAvg and FedProx are used in this process. This approach enables collaborative learning even between devices with various data distributions, significantly enhancing the learning efficiency and reliability of the entire system. Furthermore, it can strengthen the distributedness and scalability of the entire system by reducing dependence on the central server.

Table 1. Comparison of conventional AI vs. federated learning.

Category	Conventional AI Learning	Federated Learning
Data Transmission	Entire surveillance video	Parameters AI model
Security	High leak risk	Low leak risk (No transmission of Raw data)
Real-time Capability	Transmission delay	Low delay due to on-site train
Network Dependence	Needs stable network	Works with unstable networks
Operational Adaptability	Dependent on Centralized system	Autonomous to the field
Energy Efficiency	Low	High

compares the general AI learning methods and federated learning in terms of data transmission, security, real-time performance, network dependency, operational adaptability, and energy efficiency.

2.4. Technical Challenges of Federated Learning

Although federated learning has various advantages, there are numerous technical challenges that must be addressed in the actual implementation process. First, because the amount and characteristics of the data existing in each device differ, learning data imbalance problems can occur, which can affect the convergence and accuracy of the entire model. Second, when the computational performance of local devices is limited, model learning may not be performed properly or learning time may increase excessively. Particularly, security camera equipment often does not have dedicated GPUs or high-performance CPUs, so processing speed can be significantly limited depending on model complexity. Third, when various devices communicate asynchronously or connections are unstable, problems with learning cycle adjustment and parameter synchronization occur. In this case, the learning results from some devices may have negative effects on overall model performance, and the learning of the entire system may be delayed when communication interruptions or errors occur. Fourth, there are also security threats where malicious participants can disrupt the entire model by transmitting manipulated parameters. This is a new form of cyber threat that shows how federated learning can be exposed to external attacks, becoming an important consideration especially in defense and public security environments.

To solve these problems, communication efficiency, model aggregation, and data heterogeneity must be considered in federated learning algorithms, and optimized protocols and robust algorithms are needed to ensure effective learning [12,15]. Additionally, security issues such as potential attacks on aggregated parameters require advanced defense mechanisms and secure aggregation techniques for privacy protection [11,16]. While federated learning provides significant advantages in terms of privacy protection and efficiency, it also presents issues such as ensuring fairness in model training and addressing ethical considerations. As federated learning continues to develop and integrate into various areas including security camera surveillance systems, the need for robust privacy protection guarantees and scalable solutions remains important [13,17].

2.5. Expected Effects of Federated Learning in Security Camera Surveillance Environments

Security camera surveillance environments have numerous cameras that are physically distributed, and each location has distinct regional characteristics with different illumination, angles, backgrounds, and purposes of use. In such environments, learning methods that can adapt to each regional situation are required rather than a single integrated model, and federated learning can effectively realize this [18]. By having each security camera equipment perform learning autonomously, learning results that reflect the characteristics of that region can be obtained, and by integrating these at the server, a robust and adaptive surveillance system is constructed overall. Particularly in places where security is crucial, such as military facilities, access control zones, and airports, it is very useful in that AI-based real-time analysis is possible without transmitting video data externally.

Additionally, because video data is not transmitted externally, high-level security can be maintained, and the network load is also minimized [19]. This enables efficient operation even in various network environments and enhances reliability in that local equipment can maintain functions independently even if failures occur in the central server. Thus, when federated learning is applied to security camera-based intelligent surveillance systems, it provides balanced solutions in terms of performance, security, and communication efficiency, and has great potential for utilization in defense and public safety fields in the future [9,20]. Furthermore, from policy and institutional perspectives, it is consistent with personal information protection and AI utilization guidelines, so validity can be secured in both institutional compatibility and practical applicability aspects.

Recent works such as those of Casalicchio et al. [21], who developed FLWB for benchmarking FL algorithms, and Liu et al. [22], who proposed a personalized FL framework for person re-identification, demonstrate the breadth of FL research. While these studies provide concrete experimental results, our work diverges by focusing on decision-making considerations via AHP analysis rather than algorithmic benchmarking. Unlike algorithmic studies that experimentally benchmark different FL approaches, this paper focuses on deriving evaluation criteria through expert-based AHP analysis. Therefore, direct performance comparisons with state-of-the-art FL methods were not conducted. Instead, we aim to complement such works by providing a structured decision-support framework for applying FL in sensitive operational contexts.

In prior work, fog computing was adopted to enhance real-time video stream analytics in urban surveillance systems [23]. While this approach improved latency and distributed processing, it did not address critical issues of data privacy and collaborative learning across distributed nodes. Our framework, in contrast, applies federated learning to prioritize decision factors in high-security surveillance environments.

Recent research has also applied federated learning for security enhancement in UAV swarms [24]. By leveraging FedAvg and SHAP analysis, the study demonstrated collaborative learning with explainable decision-making in dynamic and distributed environments. While this work focuses on UAV networks, our study extends the discussion to surveillance systems and complements such approaches by systematically identifying critical factors for FL adoption through AHP analysis.

3. Federated Learning-Based Security Camera Detection Performance Analysis

The purpose of this chapter is to quantitatively analyze the performance advantages that federated learning provides in security camera surveillance systems compared to existing centralized AI learning methods, and to evaluate practical applicability. While the existing studies have focused on demonstrating the general superiority of federated learning, this study aims to deeply explore practical performance in the specific application environment of security camera surveillance systems. In particular, we evaluate the practical contribution of federated learning based on four key performance indicators: detection accuracy, false positive rate reduction, network transmission volume, and response time, and through this, we aim to verify its applicability in actual operating environments.

3.1. Analysis Target and Method

This chapter aims to conduct a more in-depth and systematic analysis of the performance advantages that federated learning can provide in security camera surveillance systems compared to existing centralized AI learning approaches. The technical superiority and efficiency of federated learning have already been sufficiently demonstrated through various domestic and international papers and research results. Therefore, this study does not focus on simply reconfirming the general superiority of federated learning, but rather emphasizes the significance of analyzing its applicability and practical performance specifically in security camera surveillance systems based on four key elements: detection accuracy, false positive rate reduction, network transmission volume, and response time.

Furthermore, this analysis considers detailed aspects such as the types of federated learning algorithms applied in each study, learning cycles, communication optimization methods, and data imbalance handling techniques, while empirically examining the limitations of centralized learning and the solutions provided by federated learning. Beyond simple numerical comparisons, this study contemplates the experimental conditions, constraints, and limitations identified by the researchers in each study, linking these to the realistic contribution potential of federated learning-based security camera systems.

Particularly, this research plans to conduct an Analytic Hierarchy Process analysis targeting expert groups in subsequent studies, based on the core components of federated learning derived through a literature review, to determine the priority of factors that should be most importantly considered when applying federated learning in security camera surveillance environments. Through this integrated approach, policy and technical implications will be derived together, serving as materials that can contribute to future research, surveillance system design, and technology standardization efforts.

It should be noted that the performance analysis in this chapter is based on previously published experimental studies rather than original experiments conducted by the authors. This approach was intentionally adopted to ensure that the subsequent AHP analysis is grounded in validated empirical findings, thereby avoiding redundancy while strengthening practical applicability.

3.2. Analysis of the Existing Experimental Research (Literature-Based)

This section systematically organizes the achievements of the existing research where federated learning was applied to security camera surveillance systems, focusing on four key performance factors: detection accuracy, false positive rate reduction, network transmission volume, and response time. This analysis does not merely list key numerical values, but comprehensively considers various factors such as the background of each study, experimental design, data characteristics, network environment, equipment performance, algorithm settings, communication methods, and data imbalance handling methods from multiple perspectives. This analysis is based on international prestigious academic journals and public institution technical reports, and explores more deeply how the advantages of federated learning contributed to actual security camera surveillance environments. For this purpose, we reviewed the experimental conditions, application environments, and major achievements of each study together, and approached them from a comprehensive and critical perspective that included limitations raised during the experimental process and improvement measures. This approach goes beyond a simple comparison and becomes the foundation for evaluating the practical value of federated learning in the design and policy application of security camera surveillance systems.

However, this result was achieved on a controlled dataset with limited complexity, and it remains uncertain whether similar accuracy can be obtained in heterogeneous real-world surveillance environments. Although the integration of FL and LLM achieved promising accuracy, the high computational overhead raises concerns regarding deployment on resource-constrained edge devices. In summary, while federated learning consistently outperforms centralized approaches across diverse studies, most works rely on small-scale datasets, idealized synchronization assumptions, or simplified scenarios. The challenges of non-IID data, client computational overhead, and real-world scalability remain underexplored. These limitations highlight the necessity of our AHP-based analysis, which incorporates expert judgment to prioritize critical factors beyond what the existing experimental results alone can demonstrate.

3.2.1. Detection Accuracy

Federated learning is a distributed learning method where each edge node in security camera surveillance systems learns models with local data and transmits only parameters to the central server to update the global model. It provides the advantage of ensuring data privacy without transmitting sensitive video data while improving detection performance by utilizing diverse learning data.

In particular, the FL-TENB4 architecture is a representative case that implements real-time deepfake detection in security camera environments by fusing the lightweight EfficientNetB4-Lite model with TinyML technology. This system achieves high detection accuracy while maintaining low latency even on edge devices with limited computational resources, showing performance suitable for security camera environments that require real-time video analysis [25].

Federated learning has demonstrated excellent performance in various security camera detection tasks.

Table 2. Comparative analysis with Client 5 and Client 7. Adapted with permission from Ref. [26]. Copyright 2025, IEEE.

	Client = 5		Client = 7
Method	Accuracy	Precision	Accuracy	Precision
Proposed	98.70	98.69	98.73	98.73
Xception	95.12	95.00	95.34	95.32
RestNet	93.26	93.25	93.62	93.63
ShuffleNet	96.56	96.33	96.73	96.20
RestNet50	94.60	94.56	94.72	94.72
MobileNet	95.42	95.32	95.43	95.42

shows the results of comparing the performance of various federated learning methods in Client 5 and Client 7 environments [26]. In the field of violence detection, the proposed model that combines Personalized Federated Learning (PFL) algorithms with YOLO-ResNet50 fusion models achieved the highest accuracy (98.70%, 98.73%) and precision (98.69%, 98.73%) in both client environments, showing superior performance compared to the other methods of existing models. Figure 1 shows the change in model accuracy according to communication rounds. The black dotted line represents validation accuracy (Val_Acc), and the orange dotted line represents training accuracy (Train_Acc). As can be seen from the graph, both accuracies gradually improve as communication rounds progress and finally converge to a high performance of over 90%. This suggests that consistent performance can be maintained despite data heterogeneity or environmental diversity [26,27,28].

The superiority of FL-based approaches has also been verified in general object detection tasks. A study applying an FL-based YOLOv8n model to human detection tasks recorded 33.9% improved consistency and 74.8% detection accuracy compared to traditional centralized learning methods [29]. These results demonstrate that FL is a practical solution that can improve the scalability and operational efficiency of security camera systems while meeting privacy protection requirements. Furthermore, FL-based security camera systems are reported to maintain robust performance even under complex conditions in actual urban environments or various weather conditions, making them highly evaluated for real-world deployment possibilities.

3.2.2. False Positive Rate Reduction

Federated learning is a machine learning paradigm that can train models in distributed environments while ensuring data privacy, and is particularly attracting attention as an effective method for reducing false positive rates. According to recent studies, FL has been proven to significantly reduce false positive rates compared to the existing centralized models in various security fields such as network intrusion detection, smart grids, and IoT security [26,27,28,29]. When FL was applied to network-based intrusion detection systems, model updates from distributed clients were performed reliably, showing results of reducing false positive rates by up to 13%, and it was shown that high accuracy could be continuously maintained even in real-time data environments [30,31].

FL frameworks that incorporate advanced techniques such as Graph Neural Networks or Hierarchical Aggregation are achieving even more outstanding performance improvements. In malicious client detection systems, they showed results of completely eliminating false positive rates to 0%, and in model poisoning attack defense systems, they significantly reduced false positive rates to 6.4% [32,33].

In the case of FL-based security threat detection systems operating in large-scale distributed environments, as shown in

Table 3. Performance comparison of different models in threat detection. Adapted with permission from Ref. [34]. Copyright 2025, arXiv.

Model Type	Accuracy (%)	False Positive Rate (%)
Federated learning model	94.1	3.5
Federated learning + LLM	96.4	2.9
Multimodal LLM model	93.2	3.9

, they achieved the dual effect of reducing false positive rates and false negative rates by 1.8% and 2.4%, respectively, while completely ensuring data privacy [34]. The Federated learning + LLM (large language models) model achieved the best performance with 96.4% accuracy and a low false positive rate of 2.9%. This is superior results compared to the basic federated learning model (94.1% accuracy, 3.5% false positive rate) and multimodal LLM model (93.2% accuracy, 3.9% false positive rate). In particular, it shows that the approach combining federated learning and LLM can effectively reduce false positive rates while increasing threat detection accuracy. These results clearly demonstrate that federated learning is a powerful solution that effectively reduces false positive rates even in various attack scenarios and data heterogeneity (non-IID) environments, bringing substantial security performance improvements [30,32,33,34].

3.2.3. Network Transmission Volume

Federated learning is a distributed machine learning paradigm that allows multiple distributed client devices to collaboratively learn global models without transmitting raw data to a central server. However, the repetitive model parameter exchanges that occur in each round during this learning process cause a significant network transmission volume, which particularly leads to serious communication bottlenecks in large-scale deep learning models. Various efficiency techniques are being researched to solve this communication overhead problem. The main approach is client selection techniques, which is a method of selecting only clients that can provide meaningful updates from all clients to participate in communication [35].

Table 4. Communication rate (%) with varying numbers of initial clients N. Adapted with permission from Ref. [35]. Copyright 2023, ELSEVIER.

Method	Adaptive-OU	Ignore	Zero
N = 10	46.60	46.02	47.30
N = 20	47.43	47.14	47.98
N = 50	48.95	48.93	49.13

shows the impact of initial client numbers (N = 10, N = 20, N = 50) on communication rates for three different client selection methods (Adaptive-OU, Ignore, Zero). The Adaptive-OU method achieved the highest communication rate under all initial client number conditions (46.60%), which is similar to the Zero method (47.30%). On the other hand, the Ignore method showed relatively low communication rates (46.02~48.93%). Additionally, Figure 2 shows the round-by-round performance changes of three communication strategies OU (Ornstein–Uhlenbeck), Zero, and Ignore under four different threshold conditions (Threshold = 0.02, 0.03, 0.04, 0.05). Under all threshold conditions, they show a pattern of a high communication volume in the initial rounds that gradually decreases as learning progresses. In particular, as the threshold increases (0.05), the communication volume decreases more rapidly, confirming that efficient communication is achieved, and the OU strategy shows an overall stable convergence pattern.

Moreover, methods have been proposed to reduce transmission volume by decreasing the number of bits of the model parameters transmitted through parameter quantization techniques [36], and methods for selectively transmitting only some of the entire layers of the model through selective layer transmission techniques have also been proposed [37,38]. These techniques have been experimentally verified to reduce network traffic by 30–70% while maintaining model accuracy.

In edge computing environments, more sophisticated communication optimization strategies are required. Data and model parallelization techniques distribute communication load by partitioning the learning process across multiple dimensions, and client clustering techniques group clients with similar data distributions to reduce unnecessary global communication [39,40]. Additionally, update filtering techniques can effectively reduce the number of communications by excluding minor parameter changes below a threshold from transmission.

Furthermore, parameter structural characteristic utilization techniques analyze the sparsity and redundancy of neural network parameters to preemptively remove updates that are not substantially important [41]. Recently, beyond the traditional federated averaging (FedAvg) algorithm, particle swarm optimization (PSO)-based aggregation algorithms have been proposed [42]. This approach dramatically reduces communication volume by transmitting only optimized score values instead of large parameter vectors, and has been reported to simultaneously improve both communication efficiency and model accuracy, particularly in unstable network environments.

The various communication efficiency approaches described above contribute to significantly reducing the network transmission load of federated learning systems while minimizing the performance degradation of the learned models, playing a crucial role in building practical federated learning systems.

3.2.4. Response Time

Federated learning is a distributed machine learning technique that enables collaborative model training among distributed clients without sharing raw data, effectively ensuring data privacy. However, federated learning systems inherently involve iterative communication processes between clients and servers, and the resulting communication latency and computational resource constraints of each client become major bottleneck factors causing degradation in the overall system’s response time.

This latency problem is particularly severe in Internet of Things (IoT) environments. Recent research has proposed integrated approaches to address this issue. Specifically, an FL client architecture has been developed that combines local control mechanisms with transfer learning and differential privacy [43]. Figure 3 consists of two subplots and compares the latency performance of various algorithms according to data size. The left graph shows latency according to the data size of each sample, while the right graph represents latency according to the data size of the global model. The comparison algorithms include two proposed algorithms (indicated by red and blue lines with ζ = 10 and ζ = 50), the all-sensor selection method (green line), the random sensor selection method (purple line), and the fixed pruning method (fixed pruning rate 0.1, black line). According to the experimental results, it can be confirmed that the proposed algorithms maintain relatively lower latency than other methods as the data size increases, with particularly superior performance under the ζ = 10 condition. The method of selecting all sensors shows the pattern of most rapidly increasing latency as the data size increases [43].

This integrated system has been experimentally verified to reduce the response time for user authentication and control message transmission to less than 1 s while effectively solving the learning performance degradation problem caused by limited local data through transfer learning.

For performance improvement of federated learning, device selection algorithms, model parameter quantization, and wireless resource allocation techniques can be integrally applied to improve the convergence speed of FL systems by up to 87% [44,45]. Furthermore, through advanced optimization techniques such as asynchronous federated learning, hyperparameter optimization, model partitioning, and hierarchical neural architecture search, the learning completion time can be reduced by 30~70% or more even in resource-constrained environments, while simultaneously improving model personalization performance and system response time [46,47,48,49]. These research achievements demonstrate that federated learning can be practically utilized even in application areas where real-time services or low-latency requirements are critical.

3.3. Performance Comparison

This section provides a comprehensive comparison of performance between federated learning-based security camera surveillance systems and existing centralized learning systems across key indicators. The major performance factors include detection accuracy, false positive rate, network transmission volume, and response time, with the existing research results reinterpreted under standardized conditions.

First, regarding detection accuracy, federated learning-based systems maintain an accuracy of over 98% and demonstrate performance that is 3–5% higher on average than centralized learning due to local learning effects that reflect regional characteristics. The accuracy advantage of federated learning becomes particularly pronounced as the number of clients increases. Second, the false positive rate of federated learning is up to 13% points lower than centralized learning through the provision of customized models and the introduction of hierarchical aggregation techniques. This is interpreted as a result of effectively overcoming data imbalance and heterogeneity. Third, in terms of network transmission volume, when the optimization techniques of federated learning such as parameter quantization and selective layer transmission are applied, communication volume reduction effects of up to 70% are observed. This signifies that communication resource consumption is significantly lower compared to centralized learning. Fourth, regarding response time, federated learning demonstrates competitiveness in low-latency applications thanks to its distributed architecture and communication optimization. Experimental figures report authentication and control message response times of less than 1 s and learning convergence speed improvements of up to 87%.

Table 5. Performance comparison.

Indicator	Federated Learning	Centralized AI Learning
Detection Accuracy	98% or higher	93~95%
False Positive Rate Reduction	3~4%	10% or higher
Network Transmission Volume	Up to 70% reduction	Default
Response Time	Under 1 s	3 s or more

summarizes the comprehensive comparison results for each performance indicator.

Table 5 demonstrates a comparison of performance trends between federated learning and centralized AI learning, based on the findings reported in prior studies. The values reflect typical ranges observed in the literature rather than standardized or directly averaged results. It should be noted that the values in Table 5 are extracted from different studies, each with distinct datasets and experimental environments. Rather than standardizing the raw results across heterogeneous conditions, we present them here as representative examples to illustrate the general trend consistently observed in the literature: FL-based systems outperform centralized AI in terms of accuracy, false positives, communication efficiency, and latency.

3.4. Discussion: Analysis of Practical Effects of Federated Learning

Federated learning has enabled not only a change in learning methodology but also substantial performance improvements and structural innovation in security camera surveillance systems. First, in security camera surveillance environments with distinct regional characteristics, federated learning has simultaneously enhanced detection accuracy and security by generating region-specific surveillance models through local data-based learning. Second, in terms of communication efficiency, federated learning has overcome the bottleneck phenomena of conventional centralized systems through techniques such as data non-transmission architecture, selective client updates, and parameter optimization, enabling stable operation even in diverse network environments. Third, federated learning has been effective in privacy protection and detection, and defense against malicious parameter attacks from a security perspective. Robustness against cyber threats has been secured through the refinement of aggregation algorithms and the application of anomaly detection techniques. Finally, the practical effects of federated learning extend beyond simple performance metric improvements to strengthen the legitimacy of AI-based surveillance system adoption from policy and institutional perspectives, contributing to securing scalability and sustainability. Future research requires empirical data-based validation in actual operational environments and the establishment of standardized models.

While the numerical results originate from heterogeneous studies, their collective implication is consistent: FL provides structural and performance advantages across diverse environments. Our intention is not to claim universal benchmarks but to highlight this repeated directional superiority as the basis for our subsequent AHP prioritization.

4. AHP-Based Analysis of Key Elements for Effective Federated Learning

This chapter conducts an AHP analysis to derive key elements that influence performance improvement in federated learning-based security camera surveillance systems. First, we present the system configuration diagram of the proposed federated learning-based security camera surveillance system, followed by a detailed explanation of the AHP analysis methodology. We design a hierarchical structure consisting of eight evaluation elements including detection accuracy, security, and false positive rate reduction, and conduct pairwise comparison surveys targeting expert groups to quantify the relative importance of each element. Finally, based on the analysis results, we derive the priority ranking of key elements and present practical implications.

In this study, the AHP was adopted instead of other Multi-Criteria Decision-Making (MCDM) methods such as TOPSIS (Technique for Order Performance by Similarity to Ideal Solution). The primary objective of our research was not to select one ‘optimal’ alternative system but rather to identify and prioritize the relative importance of evaluation criteria when introducing federated learning. The AHP is particularly well suited for this purpose, as it allows structured pairwise comparisons and consistency checks among expert judgments, which directly yield priority weights for criteria. In contrast, TOPSIS is more appropriate when a set of alternative systems exists and must be ranked based on their relative closeness to an ideal solution. Since the focus of this work was to derive the weighting of key factors, rather than to rank or select alternatives, the AHP was considered the most appropriate approach.

Several studies have applied the AHP or MCDM to optimize surveillance camera placement [50,51], or to evaluate intrusion detection systems [52]. These studies, however, focus on spatial optimization or detection algorithms, whereas our study emphasizes the adoption of federated learning in surveillance networks. Unlike prior AHP/MCDM applications in surveillance system design, our work provides a structured decision-support framework specifically for federated learning adoption, thereby offering a novel perspective in the literature.

4.1. System Configuration of Federated Learning-Based Security Camera Surveillance System

The system configuration diagram of the federated learning-based security camera surveillance system proposed in this paper presents a practical implementation structure for simultaneously achieving object detection performance improvement and privacy protection in distributed environments. The presented configuration diagram is designed based on algorithm structures and security techniques verified in the existing federated learning literature and systematized in a form applicable to security camera networks. Federated learning is a method that performs model training without transmitting data externally from each local device, then it delivers only the learned parameters (weights) to the central server for aggregation to improve the overall model. This approach has particular advantages in security camera surveillance environments, including minimizing privacy exposure, reducing network load, and enabling customized model learning that reflects regional characteristics.

In the proposed configuration diagram (Figure 4), multiple nodes such as security camera Nodes A, B, and C collect video data in the field, train object detection models in local environments, and then extract and transmit only the parameters of those models. This provides a foundation for safely utilizing real-time data, and since the video itself remains only locally, it is advantageous from a privacy protection perspective. The central security node receives parameters transmitted from each node and aggregates them through the FedAvg (Federated Averaging) algorithm. This algorithm updates the overall model by calculating the average of the learned weights, enabling simple yet statistically stable learning. Subsequently, the integrated AI server operated at the command center generates the final integrated model based on the aggregated model, during which the Krum algorithm [53] is applied to remove parameters from anomalous nodes. Krum is a filtering technique with Byzantine fault tolerance characteristics that mitigates the problem of overall model quality degradation due to unreliable nodes. Additionally, the integrated server prevents a reverse inference of node-specific learning information through Differential Privacy (DP) techniques and strengthens the privacy protection level of the entire system [54]. This can be considered an essential security function when considering the characteristics of security camera systems operated in sensitive areas (military facilities, hospitals, airports, etc.).

This configuration diagram represents a structure that satisfies security, real-time performance, and high-performance model training for video-based object detection, while simultaneously, effectively overcoming the limitations of existing centralized AI learning methods (privacy exposure, network bottlenecks, lack of regional reflection, etc.). Therefore, the proposed system architecture has value as a practical case demonstrating the feasibility of real-world application of federated learning technology.

4.2. AHP Overview and Analysis Methodology

This section provides a detailed explanation of the analytical procedure applying the Analytic Hierarchy Process to derive key factors that influence performance improvement in federated learning-based security camera surveillance systems. The AHP is a multi-criteria decision-making technique that systematizes complex decision-making problems into hierarchical structures and quantifies expert judgments to derive relative importance (weights) among factors, and has been widely utilized in various engineering, policy, and industrial fields [55].

Federated learning systems are affected by various technical and operational variables such as network environment, data sensitivity, and node heterogeneity. While existing studies have emphasized individual factors such as detection accuracy, privacy protection, and communication efficiency, a quantitative analysis of the relative importance of these factors is necessary in complex federated learning environments. Therefore, this study aims to derive the key influencing factors necessary to demonstrate that federated learning is structurally superior in performance compared to existing AI learning through expert judgment-based AHP analysis.

4.2.1. AHP Map Construction

The first step of AHP is to design a hierarchical structure appropriate for the analysis purpose (Figure 5). In this study, we constructed a hierarchy centered on eight evaluation factors for improving the performance of federated learning-based security camera surveillance systems.

Table 6. AHP Evaluation Factors.

Evaluation Factor	Problem Resolution Contribution	AI Performance Contribution
① Detection accuracy	Improved detection accuracy through region-specific learning	Maintains high detection performance in various environments
② False positive rate reduction	Increased alarm reliability through unnecessary alarm reduction	Enhanced alarm system quality
③ Response time	Minimized data movement, reduced communication load	Maintained real-time performance, cost reduction
④ Network efficiency	Reduced central server dependency, resolved response delays	Improved real-time detection and alarm speed
⑤ Security	Prevention of data leakage and tampering	Enhanced applicability in high-security environments
⑥ Sustainability	Guaranteed stability during system scale expansion	Applicable to large-scale surveillance systems
⑦ Scalability	Reduced local equipment burden	Secured AI performance even on low-specification equipment
⑧ Resource efficiency	Secured feasibility in maintenance and cost aspects	Provided reliability and stability for long-term operation

provides descriptions of the AHP evaluation factors.

• Goal (Level 1): Derivation of key factors for performance improvement.
• Criteria (Level 2): ① Detection accuracy, ② False positive rate reduction, ③ Response time, ④ Network efficiency, ⑤ Security, ⑥ Sustainability, ⑦ Scalability, ⑧ Resource efficiency.

These factors were derived based on literature-based prior research and the federated learning system design diagram of this study, with each factor reflecting the technical advantages that federated learning has over existing centralized AI learning. The eight factors derived in this study are configured to enable a balanced evaluation of the performance of federated learning-based security camera surveillance systems across various dimensions including accuracy, sensitivity, real-time performance, efficiency, security, operational sustainability, and scalability. Each factor operates complementarily without overlapping with others, providing sufficient criteria for judging the overall quality of federated learning. Additionally, these factors are evaluated to have high practical applicability as they reflect the problems and challenges faced in actual security camera system operations. Therefore, the AHP analysis framework of this study can function as a scientific and structural decision-making tool capable of quantitatively deriving the technical feasibility and strategic priorities of federated learning implementation.

4.2.2. Construction of Pairwise Comparison Matrix

To determine the relative importance of each factor, AHP conducts pairwise comparison surveys targeting expert groups. The survey was structured to have each expert evaluate the relative importance of eight factors in combinations of two, totaling 28 items, using a 1~5 scale. When there are n experts, each pairwise comparison result is integrated through geometric mean, and the result forms an 8 × 8 pairwise comparison matrix $A=\left[a_{ij}\right]$. Here, $a_{ij}$ represents how many times the more important factor i is compared to factor j, and the matrix satisfies the following properties:

• $a_{ii}=1$: Comparison with itself is always 1.
• $a_{ji}=1/a_{ji}$: Symmetry principle.

$$\mathrm{A}=\left[\begin{array}{cccc}1& a_{12}& \dots & a_{18}\\ 1/a_{12}& 1& \dots & a_{28}\\ ⋮& ⋮& \ddots & ⋮\\ 1/a_{18}& 1/a_{28}& \dots & 1\end{array}\right]$$

(1)

This matrix systematically represents the pairwise comparison results from expert responses and serves as the foundation for subsequent calculations.

4.2.3. Matrix Normalization and Weight Vector Derivation

After calculating the sum of each column $S_j={\sum }_{i=1}^n{a}_{ij}$, a normalized matrix $N=\left[n_{ij}\right]$ is generated by dividing each element by the sum of its respective column.

$$n_{ij}={\displaystyle \frac{a_{ij}}{S_j}}$$

(2)

Next, the weight vector $\mathrm{w}=\left[w_i\right]$ is derived by calculating the average of each row. This represents the relative importance of each factor, and the total sum equals one.

$$\mathrm{w} = {\displaystyle \frac{1}{n}}\sum _{j=1}^n{n}_{ij}$$

(3)

The weight vector is used to quantify the contribution of each factor in the final decision-making process. After calculating λ through the product of the pairwise comparison matrix and the weight vector, the maximum eigenvalue is calculated using the following equation:

$${\lambda }_{max}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\displaystyle \frac{{\left(Aw\right)}_i}{w_i}}$$

(4)

This value is necessary for quantifying the consistency of expert judgments.

4.2.4. Calculation of Consistency Index (CI) and Consistency Ratio (CR)

Based on ${\lambda }_{max}$, the Consistency Index (CI) is calculated as follows:

$$\mathrm{C}\mathrm{I} = {\displaystyle \frac{{\lambda }_{max}-n}{n-1}}$$

(5)

Next, the Consistency Ratio (CR) is calculated for comparison with a random matrix:

$$\mathrm{C}\mathrm{R} = {\displaystyle \frac{CI}{RI}}$$

(6)

Here, RI is the Random Index, and, as shown in

Table 7. Index of RI.

n	1	2	3	4	5	6	7	8	9	10
RI	0.0	0.0	0.58	0.90	1.12	1.24	1.32	1.41	1.45	1.49

, when n = 8, RI = 1.41 is applied. If CR < 0.1, the judgment is considered consistent.

The final weights represent the importance of each factor and are directly utilized in security camera surveillance system design, technology investment, and policy formulation. The CI and CR values serve as evidence supporting the reliability of the analysis results, and the results are included as essential data in system design reports and policy proposals.

4.2.5. AHP Expert Response Integration Method

In the process of applying AHP analysis for an optimal design of federated learning-based security camera surveillance systems, pairwise comparison survey responses are collected from multiple experts. To combine these into a single integrated pairwise comparison matrix, the arithmetic mean method and geometric mean method are representatively used as methods for calculating representative values. These two methods differ in their combination approach and theoretical appropriateness, and can be selectively utilized depending on the analysis purpose.

The arithmetic mean method generates an integrated pairwise comparison matrix by calculating the simple average of the expert response values for each pairwise comparison item. Expressed as a formula, it is as follows:

$$A_{ij}^{*}={\displaystyle \frac{1}{N}}\sum _{r=1}^N{A}_{ij}^{\left(r\right)}$$

(7)

where

• $A_{ij}^{*}$: an element ($i,j$) of the integrated pairwise comparison matrix;
• $A_{ij}^{\left(r\right)}$: a comparison value ($i,j$) from the $r$-th expert;
• $N$: the number of experts.

This method has the advantage of being simple and intuitive in calculation, reflecting a simple average perspective of expert responses. However, since AHP pairwise comparison values have multiplicative meanings as ratio scales, the arithmetic mean method has the limitation of being theoretically somewhat inappropriate.

The geometric mean method generates an integrated pairwise comparison matrix by calculating the geometric mean of the expert response values for each pairwise comparison item. It is expressed as the following formula:

$$A_{ij}^{*}={\left(\prod _{r=1}^N{A}_{ij}^{\left(r\right)}\right)}^{{\displaystyle \frac{1}{N}}}$$

(8)

Here, the symbols are the same as in the arithmetic mean method. The geometric mean method aligns with the theoretical foundation of the AHP, where pairwise comparison values have multiplicative relationships as ratio scales, and it also contributes to stabilizing the consistency index of the analysis results. Additionally, it has the effect of relatively mitigating the influence of extreme values, enabling stable results even when there are large response differences among experts.

When integrating expert responses in an AHP analysis, the geometric mean method is theoretically more appropriate than the arithmetic mean method and is also recommended in international standard AHP application cases. Therefore, this study performed response integration using the geometric mean method.

4.3. Expert Response Collection and Analysis Results

This study applied the Analytic Hierarchy Process to identify key factors for performance improvement in federated learning-based security camera surveillance systems and to quantify relative importance among these factors. For this purpose, structured pairwise comparison surveys were conducted targeting experts in related fields, and hierarchical analysis procedures were performed based on these results.

4.3.1. Overview of Expert Response Collection

Survey responses for AHP analysis were collected from a total of 15 experts. The panel included researchers and practitioners from academia, defense research institutes, and industry. Their fields of expertise covered AI algorithm design (five experts), video surveillance and computer vision (four experts), cybersecurity and data privacy (three experts), and defense/military surveillance operations (three experts). All participants had more than 10 years of professional experience, and several currently serve as professors, senior researchers, or system architects. Experts were selected based on their direct involvement in federated learning, AI-based surveillance, or high-security monitoring systems. To preserve anonymity, only aggregated information is reported. To ensure the expertise and validity of survey responses, the following criteria were applied:

• Personnel with practical experience in security camera systems, perimeter surveillance, and related fields;
• Individuals with academic knowledge related to federated learning AI algorithms;
• Personnel with both types of experience.

The survey was provided in an online pairwise comparison format, and each expert was requested to evaluate the importance of all possible pairwise comparisons (28 pairs total) among eight major evaluation factors using a 1~5 scale. Additionally, to ensure response consistency, a survey platform designed to automatically calculate the consistency index and consistency ratio during response was used.

4.3.2. Response Integration and Consistency Result

The pairwise comparison matrices derived from individual expert responses were first integrated using the Geometric Mean Aggregation method. Relative importance vectors (weights) were calculated through eigenvectors from the aggregated group average matrix. For the 8 × 8 pairwise comparison matrix constructed by collecting expert response data, after normalizing each matrix and calculating weight vectors, the maximum eigenvalue ${\mathrm{\lambda }}_{max}$ was obtained as 8.014. The CI was calculated as 0.002 using the following formula:

$$\mathrm{C}\mathrm{I} ={\displaystyle \frac{{\lambda }_{max}-n}{n-1}}= {\displaystyle \frac{8.014 - 8}{8 - 1}}={\displaystyle \frac{0.014}{7}}\approx 0.002$$

(9)

The random index was set to 1.41 based on the eight-factor criteria according to Table 7. Using CI (0.002) and RI (1.41), CR was calculated as 0.001 using the following formula:

$$\mathrm{C}\mathrm{R} = {\displaystyle \frac{CI}{RI}}={\displaystyle \frac{0.002}{1.41}}\approx 0.001$$

(10)

Figure 6 represents the comprehensive AHP analysis results for evaluating key factors in federated learning-based security camera surveillance systems. The 8 × 8 normalized pairwise comparison matrix was constructed through geometric mean aggregation of expert responses, with each element representing the relative importance weights derived from the hierarchical analysis process.

The resulting CR of the derived matrix was 0.001, which satisfies the CR < 0.1 criterion. Such a low value indicates that expert judgments were highly consistent, likely due to the shared professional background and domain-specific expertise of the panel. While this increases confidence in the internal consistency of the results, it may also reflect a limited variability in perspectives. Therefore, the findings should be interpreted as reflecting a strong expert consensus rather than diverse, potentially conflicting opinions. In particular, ${\lambda }_{max}$ was derived at a level slightly exceeding eight, which theoretically represents perfect consistency, and the corresponding CI value also showed relatively low figures, suggesting that experts made relatively logical and consistent judgments regarding the relative importance among individual items.

Additionally, although minor deviations were observed in some expert responses during the analysis process, the reliability of the overall system analysis was secured by constructing a stable pairwise comparison matrix based on majority opinion through the average combination method (geometric mean application). In the future, consistency could be further improved by increasing the number of survey respondents to perform repeated measurements or by identifying respondents with large deviations in advance and requesting re-surveys.

4.3.3. AHP Analysis Results

The final weight vector quantifies the relative importance of each evaluation factor contributing to performance improvement in federated learning-based security camera surveillance systems. The AHP analysis results yielded the following importance ranking (

Table 8. AHP analysis results (importance ranking).

Evaluation Factor	Relative Importance (Weight)	Priority Ranking
Detection Accuracy	0.203	1
False Positive Rate	0.167	2
Security	0.121	3
Network Efficiency	0.113	4
Response Time	0.110	5
Sustainability	0.106	6
Scalability	0.095	7
Resource Efficiency	0.085	8

and Figure 7):

The analysis results reveal that detection accuracy emerged as the most critical factor, indicating that it serves as a key indicator for assessing the high-level object identification and anomaly detection that federated learning can achieve in actual surveillance environments compared to existing AI learning methods. Subsequently, false positive rate reduction and security occupied high importance levels, demonstrating that the structural characteristics of federated learning—reducing unnecessary alarms and preventing original data from being leaked externally—are highly valued in security-sensitive environments. Conversely, scalability and resource efficiency showed relatively lower weights, indicating that structural and operational factors were somewhat deprioritized compared to technical factors that directly impact system performance.

4.4. Key Factor Derivation and Priority Analysis

This AHP analysis quantifies the knowledge and judgment of expert groups, providing practical design priorities for performance improvement in federated learning-based security camera surveillance systems. Particularly for agencies and practitioners seeking to introduce AI-based surveillance technology in high-risk areas, national defense, and critical infrastructure, the following strategic implications can be presented:

• For performance optimization, initial deployment should focus on designing algorithms and system structures centered on detection accuracy (false positive rate reduction) and security.
• Resource efficiency and network efficiency can be important considerations in hardware-constrained environments, with the possibility of increased relative importance particularly in edge device-based operating environments.
• Scalability and sustainability can be emphasized primarily during long-term operation and expansion phases rather than initial deployment.

The diversity and expertise of the 15-member expert panel enhance the validity of the AHP results, ensuring that the derived priorities reflect both technical and operational perspectives relevant to real-world deployment. The extremely low CR value (0.001) demonstrates a strong alignment of expert opinions. However, it also highlights a limitation: the relatively homogeneous expertise of the panel may have reduced diversity in judgment. Future studies could expand the expert pool to include more varied backgrounds to test the robustness of these priorities.

In conclusion, the AHP results of this study present empirical design criteria that consider the balance between functional performance and security requirements in federated learning technology adoption. These findings are expected to serve as valuable reference materials for future policy formulation, system design, and operational strategy development.

5. Conclusions

This study confirmed through theoretical analysis and empirical comparison that introducing federated learning to security camera surveillance and perimeter systems can provide various advantages in terms of privacy protection, communication efficiency, security, and regional specialized modeling compared to existing centralized artificial intelligence learning approaches. Centralized AI requires moving data collected from numerous security camera equipment to servers, which causes problems such as privacy violations, network delays, and increased communication costs, and these limitations become even more pronounced in actual environments. In contrast, federated learning has a structure where learning occurs individually while data remains within local devices, and only the learned weights are transmitted to the center, making it possible to fundamentally reduce the risk of data leakage while enabling efficient AI learning.

Based on prior research and performance experiment results, federated learning-based security camera object detection systems showed competitive results in terms of learning speed, transmission efficiency, and scalability along with better detection accuracy compared to existing methods. Particularly in surveillance environments where various environments and conditions coexist, it was confirmed that the distributed learning structure provided by federated learning can be a more practical solution. Furthermore, this study conducted an Analytic Hierarchy Process analysis targeting expert groups to identify key factors for an effective introduction of federated learning. The analysis results showed that security assurance, communication resource reduction, local processing performance for reflecting regional characteristics, and system scalability had high importance, which can be utilized as criteria to consider when designing federated learning systems in the future. Therefore, rather than focusing on the performance advantages of a specific algorithm, this paper contributes by providing guidance on what should be prioritized at the policy, technical, and operational levels when adopting federated learning.

In conclusion, this study presented that federated learning is a powerful and realistic alternative that can replace existing artificial intelligence learning structures in security camera surveillance environments, and proposed directions for expanding its application scope through empirical application and technological advancement in actual environments in the future.