1 Introduction

The Internet of Things (IoT) technology, which has applications in vital infrastructure, has revolutionized the creation of intelligent information systems. These systems improve the effectiveness of technology and business process management by automating tasks like equipment monitoring, lifecycle management, energy consumption control, and facilitating the best linkages between enterprises and service users [1]. Various industries, including smart grids, transportation systems, smart cities, and smart homes, are adopting IoT-based services, applications, and information systems. But along with its many advantages, the growth of IoT also raises new issues and worries about information security. Due to the diversity of connected devices with varying capabilities and communication protocols, the Internet of Things (IoT) introduces intricate security challenges, amplifying privacy concerns amidst the massive data influx. The shift to edge computing necessitates a reevaluation of security models. IoT devices present challenges for timely updates because of their extended lifespan and deployment in exposed environments, which necessitate heightened attention to physical security. A comprehensive cybersecurity strategy that includes secure device design, encryption, access controls, and stakeholder collaboration is required to address these complexities. It is essential to establish a robust security framework to effectively navigate the constantly changing IoT landscape. IoT devices, constrained by processing limitations and inadequate security, are highly vulnerable due to widespread interconnection. Ensuring IoT security requires prioritizing device-level protection, implementing encryption, and skillfully managing interconnected systems to minimize risks. The increasing number of networked gadgets raises many possible concerns that could endanger people’s privacy and general security [2].

Considering the data in reference [3], it is possible to divide the security concerns connected to any IoT system into four major classes. These risk categories cover risks relating to device and user authentication, physical effects on system components, breaches of integrity, and availability, confidentiality, and risks relating to the handling of sensitive data, such as personally identifiable information. Personal data security and other sensitive information security are more critical than ever in the current environment [4]. This includes many categories of data, including biometric data, health records, and other kinds of information. Some components fall under the definition of personal data stated by the General Data Protection Regulation of the European Union [5]. Using these components to locate the people they belong to is possible. As a result, information security management systems frequently use these figures. Because of this, processing personal data under information security management obligations in IoT-based systems is more dangerous [6]. In addition, any unauthorized access to the information security management system puts at risk the privacy of user data and the security of essential information for IoT devices [7].

Cybercriminals use security system flaws to access confidential information, obstruct services, stop operations, steal data, and damage computer systems. In addition to having a negative impact on industrial production, these illicit acts continually endanger national security. The seriousness of such calamities was recently highlighted by a significant Distributed Denial of Service attack in Ukraine that caused substantial disruptions and raised public concern. In many data network contexts, accurate network attack detection can aid in locating security issues and enhancing security defense. Network attack detection is becoming a more widespread activity in considerable data traffic. A comprehensive strategy that includes packet filtering, behavioral analysis, deep packet inspection, honeypots, deception technology, NetFlow analysis, SIEM, threat intelligence feeds, real-time incident response systems, encryption analysis, and signature-based and anomaly-based detection is needed to detect network attacks in data traffic. To provide a secure environment for all users, a comprehensive strategy concentrates on ongoing monitoring, updates, and a proactive response plan for strong network security. These systems monitor network activity to identify potential threats. Signature-based IDS relies on predefined patterns, while anomaly-based IDS detects deviations from normal behavior. Machine learning enhances threat detection by analyzing user and network entity behavior for anomalies and attack patterns. Thus, it is crucial to create and use cutting-edge computational techniques developed expressly for identifying network risks in large-scale data settings [8].

The traditional ways for recognizing network assaults, such as standard statistical approaches, behavioral techniques, and Deep Packet Inspection (DPI), present major computational hurdles in big data application scenarios with a vast scale, numerous devices, and massive traffic. These methods demand message parsing and constant monitoring of the health of all networks and devices [9]. But in the domains of deep learning, evolutionary computing, and machine learning, where the emphasis is on creating behavior-based intrusion detection systems, tackling this problem has emerged as a critical issue. It is now possible to automatically identify the behavioral traits of network traffic and the underlying patterns by utilizing these cutting-edge technologies [10, 11]. Based on the characteristics and laws they have learned, they may classify the behaviors in anomalous traffic. Deep learning techniques employ various neural network models to fit highly complex nonlinear functions. Deep learning captures complex nonlinear functions using a variety of neural network models, including FNN, MLP, CNN, and RNN. While RBFN employs radial basis functions for interpolation, GAN generates synthetic data, and autoencoders concentrate on unsupervised learning. Transformers, originally designed for natural language processing, exhibit adaptability through self-attention mechanisms. These models support task-specific adaptation and are updated frequently to reflect new deep-learning architectures and advancements. As a result, they can accurately recognize the complexities of network attack patterns.

A fresh approach to distributed machine learning is federated machine learning. The fact that learning is done locally or on data collection devices is a key component of this strategy. As a result, local models are developed first; then, they are combined to create a global model. This federated learning capacity makes the development of systems that safeguard the confidentiality of sensitive data and personal information possible. A host intrusion detection system (host IDS) is one of these systems. It can monitor and analyze data obtained from and related to a host to find intrusions or other policy violations on that host. It is usually placed on the target host, although some of its capabilities can be centralized if host information is available. Unfortunately, this centralization drastically restricts its capabilities because only a few functions (such as gathering and analyzing system logs) can be done thus without risk. As a result, it can be strictly essential to deploy a host IDS on the target host based on the features and capabilities requested. These devices’ integration with embedded industrial machinery is challenging because of their constrained processing capabilities and strict operational constraints.

The following is the paper’s primary contribution:

  • To address this issue, this article provides two privacy assessment metrics for FL-based HIDSs, including a privacy score that reflects how similar the original and restored traffic attributes are.

  • The real-time model was built using the CICIDS2017 dataset, which includes numerous recent attacks.

  • An adaptive threshold-correlation algorithm (ATCA) is also introduced to improve detection accuracy by dynamically modifying threshold values based on traffic patterns and intrusion behaviors.

  • The FL-HIDS framework is created and tested using a realistic network dataset. Federated Learning focuses on privacy and security in training models on decentralized devices with sensitive data. It requires efficient aggregation and secure communication protocols. Choosing the right architecture for host-based intrusion detection systems (HIDS) on edge devices is crucial in federated learning.

  • The federated learning technique successfully uses network device collective intelligence, allowing for continuous learning and adaptation to evolving attack strategies.

This article’s remaining sections are organized as surveys: the second section examines and comments on some of the essential connected works. Section 3 explains the proposed strategy. We conduct an empirical analysis and experimental setup in Sect. 4. Finally, Sect. 5 brings this essay to a close.

2 Literature Survey

To detect such attacks, the approach proposed by Khatri et al. [12] examines the transmitted packet properties, judges the models, and visualizes the outcomes. A dataset with more than 80 columns and 10,48,575 rows is used in this investigation. Several machine learning algorithms, including random forest (RF), decision tree (DT), Ada boost classifier (ADA), ridge classifier (ridge), logistic regression (LR), SVM-linear kernel (SVM), naive Bayes (NB), and quadratic discriminant analysis (QDA), were used in the process of building and testing the models. The results reveal an exceptional accuracy rate. When autoML was used, it generated a best-fit algorithm with 99% accuracy and produced specific metrics for each data file. Next, a prediction was made using these saves, and an output CSV was created. The Autoviz package was then used to map the CSV into several displays.

Tang et al. [13] provide a federated learning-based network intrusion detection method in their study. This ground-breaking method enables several ISPs or businesses to collaborate on deep learning training while safeguarding the privacy of their local data. Doing this increases the model’s detection accuracy while preserving network communication’s anonymity. In this study’s trials, network intrusion detection data were collected using the CICIDS2017 system. These studies’ findings show that federated learning participants acquire higher detection accuracy. Furthermore, federated learning exhibits performance equivalent to centralized deep learning models when accuracy and other metrics are considered.

In their study, Zczepanik et al. [14] revealed the core ideas of a heuristic algorithm that uses a neural network. Before implementing the machine learning phase, they used a specific dataset and offered details on data processing and any necessary adjustments. It demonstrated how to modify the hyperparameters of artificial neural networks and how they learn. The models’ accuracy of more than 98% and their F2-score of more than 95% showed how precise and highly effective this method was in the results. It makes intrusion detection incredibly effective.

According to Jeune et al. [15], internet technology has advanced significantly. Security concerns increased as the number of users increased. The organization must ensure data security. Network traffic packets are examined by anomaly-based intrusion detection systems (IDS) to look for assaults. Threats were carefully identified using a neural network with an indicator variable utilising rough set for attribute reduction. This approach identifies potential dangers by carefully examining the attributes and applying a rough set theory.

A thorough analysis of network intrusion detection datasets is presented by Ring et al. [16], focusing on 15 distinct features. Their work offers a detailed analysis of these characteristics and how they apply to network intrusion detection. As they provide both a condensed overview and a substantial in-depth table and discussion, their work can be used as a guide when choosing appropriate public datasets for a specific objective. Finally, they arrive at a few conclusions that apply to all future studies that use NIDS datasets. For instance, they talk about how unlikely it is to have a flawless dataset and suggest combining many datasets for evaluation. Thus, their examination does produce helpful information even when it does not offer insight into the performance of specific algorithms or techniques.

A comprehensive intrusion detection system (IDS) was investigated by Ferrag et al. [17] in their work to combat distributed denial-of-service (DDoS) attacks. The suggested IDS employed three alternative models: convolutional neural networks, deep neural networks, and recurrent neural networks. Using the two brand-new real-world traffic datasets, TON_IoT, and CIC-DDoS2019 the researchers assessed the performance of these models. These datasets covered a wide range of DDoS attacks, enabling a thorough evaluation of the model’s performance for the binary and multiclass classification types.

Aashmi et al. [18] have developed a novel intrusion detection system that takes advantage of federated learning in the context of jungle computing. Scalability, availability, fault tolerance, and the ability to handle many computing paradigms simultaneously are just some of the benefits that jungle computing provides. We add meta-tags to the job descriptions to highlight frequently completed jobs. This real computing jungle’s code is built to function effectively on various hardware, including desktop and laptop computers, clusters, grids, the cloud, low-cost gadgets, and cell phones. The computer system architecture of the Jungle can easily be changed by just changing the meta-tags given to jobs. Frameworks for federated learning with privacy enhancements are more resistant to models of anomaly detection. The federated learning architecture is vulnerable to attacks from hostile parties. Hence, a more resilient paradigm was used in jungle computing by using Java parallel processing in various application scenarios.

Liu and Shi [19] first suggested a hybrid intrusion detection system (IDS) using genetic algorithms (GA). The results of a random forest (RF) classifier were used to generate a new fitness function. In each iteration, the chromosome with the highest value replaced the one with the lowest value. According to test results, the model consistently achieves more than 90% accuracy when applied to the reference datasets NSL-KDD and UNSW-NB15.

The benchmark datasets Bot-IoT, UNSW-NB15, and CIRA-CIC-DOHBrw-2020 were used by Halim et al. [20] to investigate their suggested GA-based feature selection approach (FBFS). The fitness function chooses the characteristics with the fewest correlations for categorization in their experiment using unsupervised GA. Their recommended method, which combines the three classification models XGBoost, k-nearest neighbor (KNN), and support vector machine (SVM), has a 99.80% success rate.

2.1 Limitations of Existing System

  • Federated learning relies on several parties or hosts contributing local data for model training. However, acquiring real-time traffic data for host intrusion detection can be difficult. Because not all hosts are ready to reveal sensitive traffic information, data availability may be limited, and models may be biased.

  • Frequent communication between the central server and participant hosts is required in federated learning. Real-time traffic monitoring necessitates low-latency processing, and the increased connection overhead might cause delays, reducing the system’s overall performance and responsiveness.

  • Hosts’ hardware capabilities, network circumstances, and data distributions might differ dramatically in a federated learning environment. This heterogeneity can make reaching uniform convergence during model training difficult. Different hosts may have varying levels of accuracy, resulting in potential variances in intrusion detection system performance.

  • Federated learning seeks to protect the confidentiality of regional data stored on participating hosts. Since real-time traffic analysis involves sensitive information regarding network activity, protecting data security and privacy is critical. Data integrity must be protected, and potential assaults on the federated learning process, such as model poisoning or inference attacks, must be avoided.

  • Real-time traffic analysis frequently deals with a high volume of data supplied by multiple hosts simultaneously. Scaling federated learning to accommodate an increasing number of servers while processing enormous amounts of traffic data can be difficult. The system must manage growing computing and communication demands while being

2.2 Problem Identification of Existing System

  • Traditional methods for detecting host intrusions in real-time traffic analysis frequently rely on centralized data collection, in which all network traffic data is provided to a central server for analysis. However, this strategy has significant drawbacks, including privacy concerns, bandwidth constraints, and the possibility of single points of failure. Federated learning addresses these challenges by allowing collaborative analysis while keeping data dispersed. On the other hand, implementing federated learning for host intrusion detection in real-time traffic analysis presents its own set of obstacles.

  • In real-time traffic analysis, networks can be highly heterogeneous, containing many devices, operating systems, and network topologies. This variety presents a difficulty for federated learning since the models must be trained in various network conditions while assuring compatibility and accuracy across all participating servers. For federated learning to be most successful and efficient in host intrusion detection, network heterogeneity problems must be identified and fixed.

  • Hosts involved in federated learning for intrusion detection may have varied computational capabilities and resource constraints. Some hosts may need more processing capacity, memory, or energy, limiting their ability to actively engage in the learning process. Designing federated learning algorithms that can accept such resource-constrained hosts is a considerable problem while retaining the system’s overall efficiency and accuracy.

  • Host intrusion detection involves sensitive network traffic data and potential security vulnerabilities. Data remains on the hosts with federated learning, decreasing privacy risks. However, maintaining data privacy and security while aggregating and updating models across dispersed servers becomes crucial. Potential vulnerabilities must be addressed, such as model poisoning attacks, information leakage during communication, or malicious behavior from collaborating hosts.

  • Real-time traffic analysis necessitates the discovery and response to intrusions on time. Due to the necessity for communication, coordination, and aggregation of model changes over numerous hosts, federated learning imposes a significant delay. Balancing the trade-off between real-time analysis capabilities and precise intrusion detection via federated learning is challenging.

3 Proposed System

In host intrusion detection systems (HIDS) that seek odd behavior, deep learning (DL) approaches are often employed to discover damaging data. Deep learning strategies for detecting damaged data have included autoencoders, variational autoencoders (VAEs), generative adversarial networks (GANs), outlier detection models such as Isolation Forests and one-class SVM, residual analysis, and data augmentation and denoising techniques. These techniques assess reconstruction errors, generate synthetic samples, detect deviations from normal data distribution, and improve model resilience to damage. The overall efficacy of identifying damaged data can be increased by combining techniques or using hybrid approaches. Autoencoders, along with anomaly detection techniques like variational autoencoders and outlier detection algorithms, are extensively used. Autoencoders reconstruct input data, while anomaly detection models identify deviations from normal patterns, indicating potential data damage. The federated learning (FL) architecture makes DL-based methods more valuable, which protects user data while training a global model. By decentralizing the training process, the federated learning (FL) architecture revolutionizes deep learning (DL) methods, allowing models to be trained on local data without threatening user privacy. FL iteratively improves a global model by sending model updates rather than raw data to a central server, ensuring that sensitive user information remains secure on individual devices. This approach not only protects user data but also increases the value of DL-based methods, particularly in applications where privacy is important. FL strikes a balance between the power of deep learning and the need to protect individual privacy by enabling collaborative learning across distributed devices, ultimately establishing a global model that has learned from diverse, decentralized data sources. However, FL-based HIDSs’ privacy protection has yet to be fully assessed. To tackle this challenge, we offer two privacy assessment metrics for FL-based HIDSs, including a privacy score that rates how close the original and restored traffic attributes are. The real-time model was created using the CICIDS2017 dataset, which includes numerous recent attacks. An Adaptive Threshold- correlation Algorithm (ATCA) dynamically adjusts threshold values based on traffic patterns and incursion behavior to improve detection accuracy. ATCA is designed to enhance detection accuracy by dynamically adjusting threshold values based on real-time observations of traffic patterns and incursion behavior. The process involves continuous monitoring, initial threshold setting, real-time analysis, dynamic threshold adjustment, correlation with known anomalies, and a feedback loop for continuous improvement. By adapting to changing conditions and learning from detections, ATCA provides a flexible and responsive approach to identifying both known and emerging threats in a given system. A realistic network dataset is used to design and test FL-HIDS. Experimental results show that the suggested technique is more precise and scalable than standard intrusion detection systems. The federated learning strategy uses network devices’ collective intelligence to learn and adapt to emerging attack strategies continuously.

Federated learning-host intrusion detection systems (FL-HIDS) development and testing necessitate careful consideration of network topology, communication latency, security, privacy of data, and diversity of hosts and data. It is critical to ensure efficient model synchronization as well as manage network topology and communication latency. The variety of hosts and data, including operating systems, applications, and network behaviors, makes robust intrusion detection difficult. For accurate and resilient FL-HIDS, decisions on model architecture, hyperparameters, and handling adversarial attacks are critical. Other critical factors are data distribution, imbalance, scalability, and regulatory standard compliance. A realistic network dataset, which includes a variety of traffic patterns and attack scenarios, is required for training and testing. The FL-HIDS plan is depicted in a block diagram in Fig. 1. Users give the data firewall information. A firewall is a constant filter for data attempting to enter a network, screening it for potential threats and blocking those that do. Before reaching the firewall, the attacker attacks the user data. This prevents data from being written to a dataset. Data is gathered from a dataset. Techniques for pre-processing collected data. Data transformation and pre-processing to remove outliers and unnecessary instances. The datasets used contain values that are symbolic, continuous, and binary. The deployment of federated learning for host intrusion detection calls for sharing privacy concerns among organizations to safeguard user privacy and expand the knowledge base of learning models. Host intrusion detection requires analyzing sensitive data, emphasizing the importance of collaborative discussions among organizations. Open dialogue fosters trust, transparency, and a commitment to ethical practices, aiding in the identification and strategy development to minimize privacy-related incidents. Introducing a more extensive variety of attacks and eliminating any attack variants within the learning model is crucial to guaranteeing the accurate detection accuracy for previously unrecognized traffic within an organization. It boosts the model’s robustness, enhancing its capability to recognize and defend against a broader range of threats. This strengthens overall security and adaptability, resulting in more effective and reliable performance across diverse scenarios. The Adaptive Threshold-correlation Algorithm (ATCA) can improve detection accuracy by dynamically altering threshold levels based on traffic patterns and intrusion behaviors if an attack is detected. Without any attacks, the information is transmitted to the server.

Fig. 1
figure 1

Proposed method of FL-HIDS

Full size image

3.1 Dataset Description

The Canadian Institute for cyber security debuted the CICIDS 2017 in 2017 [21]. The dataset includes several benign attacks that are further divided into 15 subcategories and seven primary categories of recent attacks [22]. Two million eight hundred thirty thousand one hundred eight occurrences total in the sample, with 83.3% benign and 16.7% malicious traffic. Most of the most recent real-world attacks are included in this dataset, which has been rationalized. From a five-day data collection campaign from July 3 to July 7, 2017, a program called CICFlowmeter was used to extract various flow-based, time-based, content-based, host-specific, and additional auxiliary features and their labels. Some attacks include brute force SSH, brute force FTP, Heartbleed, DoS, infiltration, botnet, web assault, and DDoS. The abstract behavior of 25 people was built utilizing a range of network protocols, including FTP, HTTPS, SSH, HTTP, and email protocols, to give the dataset more realism. Several CSVs are also included in the dataset for usage in deep learning and machine learning applications, along with the entire packet payload in PCAP format for real-time implementation. Table 1 has a thorough list of complete signatures for several attacks and the relevant types.

Table 1 The CICIDS2017 dataset provides descriptions of numerous attacks and their kinds
Full size table

3.2 Data Pre-processing

In the world of data mining, the preprocessing phase of the data is both significant and time-consuming. Dealing with accurate data frequently gathered from various systems that may be unpredictable, redundant, incomplete, or inconsistent is a requirement. Accurate data with unexpected patterns or outliers can lead to unpredictable model behavior. Redundant data increases the risk of overfitting and reduces generalization. Incomplete data introduces bias, hindering pattern extraction and impacting predictions. It is essential to transform raw data into a format appropriate for analysis and information discovery. The transformation of raw data for analysis is crucial due to the inherent complexity and lack of structure in raw data, often containing errors and inconsistencies. This process involves cleaning and pre-processing to enhance data quality and enable integration from various sources. Normalization and standardization play a role in achieving consistent presentation, while feature engineering enriches the dataset. Dimensionality reduction simplifies the data, enhancing analytical techniques and models. Then, the transformed data aligns with governance standards, assists compliance efforts, and facilitates exploration, visualization, and interpretation. Data manipulation and eliminating redundant and outlier events were part of the preparation procedure in this investigation. Symbolic, continuous, and binary values can be found in the datasets [23]. Data preparation involves transforming raw data for effective processing and analysis. Key steps include collecting, cleaning, and labeling data for machine learning algorithms, followed by exploration and visualization to ensure suitability for analysis.

3.3 Federated Learning for Host Intrusion Detection

Organizations should employ a federated learning strategy to share information about privacy issues to broaden the information base of learning models and prioritize user privacy. The learning model is exposed to various benign and attack variations to accomplish consistent detection accuracy through beforehand undetected organizational traffic. This exposure gives the model a chance to become accustomed to a broader variety of good and bad events, improving adaptability and ensuring consistent performance. Federated learning is a machine learning technique that allows for model training on decentralized devices or servers while maintaining data privacy by keeping data local. It enables diverse data sources, reduced latency through edge computing, and adaptability to dynamic environments in real-time traffic analysis. Federated learning enables local threat detection, privacy-preserving collaboration, adaptive threat models, reduced false positives, and scalability through a distributed architecture for intrusion detection. By fostering robustness and resilience, this approach enhances the model’s ability to generalize across diverse situations. It enables the model to learn effectively from both successful and challenging experiences, equipping it to handle real-world complexities, unforeseen scenarios, and noisy data. The proposed architecture facilitates collaboration between organizations through the sharing of cyber intelligence and insights. Additionally, enterprises that cannot now amass and keep enough network traffic for training learning models by partnering with other businesses can create efficient deep learning (DL) systems. Our method tackles the issue of data scarcity and enables the creation of DL-based HIDS without the need for a sizable training data set. Each member provides a small number of data samples, which make up the system’s success. The framework for federated learning is shown in Fig. 2.

Fig. 2
figure 2

Federated learning’s organizational structure

Full size image

Furthermore, federated learning assures the confidentiality and accuracy of sensitive individuals’ network information by distributing local network data samples across numerous companies. A central server oversees and manages the autonomous local model training in this federated learning setup. Despite being housed inside one of the partner organizations, the project’s global server is designed to be hosted externally by a responsible third party, such as cloud computing. Selecting an external hosting solution for the global server in federated learning depends on scalability and operational efficiency. Prioritize strong encryption, strict access controls, regulatory compliance, and thorough vetting of hosting providers to safeguard sensitive data. Each member organization must consistently log its local network data traffic according to the framework’s primary requirements. References [24, 25] provide a comprehensive list of advantages to using a uniform feature set. Federated learning is more effective since the global model may identify significant patterns within a predetermined range of data attributes. The global model’s parameters and organizational structure were created to comply with the industry-recognized network logging format.

The whole procedure is distinct by Algorithm 1, where P is the member organizations indexed by p, w is the initialized regular of limitations, t is the federated learning round, and m is the average learning rate across all member organizations. The letters E stand for the local learning rate, the regional training set, the number of local epochs, the prediction loss (xi, yi), and the local prediction loss. The regional training batch’s size is B., similar to traditional federated learning approaches;

Step 1 An ML model with a set architecture and parameters is launched by a global server to begin the process.

Step 2 To each participant, the model is sent.

Step 3 The framework is refined and trained nearby using internal network data examples.

Step 4 The weights are modified before returning to the primary server.

Step 5 The system employed is FedAvg [26], in which the server pools the weights supplied by other organizations to create a more effective intrusion detection model. This updated model considers the characteristics of each participant’s network using a more accurate set of variables.

The FedAvg approach is defined as

$$ v_{t + 1} \leftarrow \sum\limits_{l = 1}^{L} {\frac{{m_{l} }}{m}} v_{t + 1}^{l} $$
(1)

To enhance detection performance in all network scenarios, these five phases can be repeated to build a single federated learning round.

figure g

This work broadens the application of federated learning by treating each local client as an autonomous entity with a distinct selection of varied data examples. Creating a robust information security program is crucial for businesses to minimize the impact of data breaches, safeguard data privacy, and detect/prevent threats. This includes managing authorized user access, maintaining integrity with approved changes, and ensuring confidentiality through effective access control. The direct result is the creation of a comprehensive DL-based HIDS through inter-organizational collaboration without the requirement to disclose participant data to protect participant privacy. The final model’s capacity to identify threats from a more extensive range of sources is essential for an organizational defensive system. The defensive system is an important component of organizational security, which protects IT infrastructure from unauthorized access and threats using measures such as access controls and encryption. It incorporates firewalls and intrusion prevention systems to monitor and control network traffic, while regular assessments, incident response planning, and user awareness initiatives improve its effectiveness. Integration with intrusion detection systems enables timely detection and response to malicious activities. The system’s continuous improvement, adherence to compliance, and adaptability to emerging threats all contribute to a resilient security posture, making it critical for proactive intrusion detection and prevention. Creating a strong information security program for businesses is vital. It helps identify and prevent risks, protects data privacy, and minimizes the impact of potential data breaches. This includes ensuring confidentiality through access control, maintaining integrity with authorized changes, and ensuring availability for authorized users. As a result, a solid learning model is developed that can discern between good and bad heterogeneous traffic thanks to its extensive knowledge and insights. When there are modifications to the regular traffic patterns due to upgrades to the Standard Operating Environments (SOE), sophisticated models like these can potentially reduce the incidence of false warnings. Federated learning allows personalized model adjustments on local devices, reducing centralized data transfer. This improves accuracy in anomaly detection by adapting to local conditions, enhances privacy by keeping data localized, and minimizes false warnings through collaborative learning in dynamic environments. This improvement is because the models have become more adaptable to variations in benign traffic distribution due to learning from secure usage across multiple networks. Extracting harmful patterns from a broader range of attacks that target numerous organizational networks also promotes a better incidence of sophisticated and zero-day attack detection.

3.4 Adaptive Threshold-Correlation Algorithm (ATCA)

Look at Fig. 3’s general nonlinear function first.

Fig. 3
figure 3

Adaptive “threshold” parameter for a nonlinear function

Full size image

The function \(f(y;B)\) must be peculiar and limited. \(f(y;B)\) can change its shape by inserting a parameter B known as the “threshold,” as seen in the picture.

Assume f(x; A) is written as

$$ f(y;B) = B.\widetilde{f}(y/B) $$
(2)

where \(\widetilde{f}()\) is a fundamental function with attributes.

$$ {\text{(a}})\;\;\widetilde{f}(u) \to u\,as\,|u| \to 0; $$
(3)
$$ ({\text{b}})\;\;\;\widetilde{f}(u) \to A.{\text{sgn}} (u)\,as\,|u| \to \infty ,\,\,with\,\,A \ge 0 $$
(4)

It is worth noting that \(f(y:1)\) equivalents are the basic function \(\widetilde{f}(y)\).

The nonlinear correlation function, designated as f(x; A), is incorporated into the adaptation procedure’s implementation. The following is a presentation of the equation for updating the tap weights.

$$ d^{(n + 1)} = d^{(n)} + \alpha_{c} f(e_{n} + v_{n} ;B^{(n)} )b^{(n)} $$
(5)

where \(b^{(n)} = [b_{n} ,b_{n - 1} ,....b_{n - N + 1} ]^{T}\) stands for the tap weight vector at time n (T for transposition), \(d^{(n)} = [d_{0}^{(n)} ,d_{1}^{(n)} ,....d_{N - 1}^{(n)} ]^{T}\) for the tap input reference signal vector at time n, en for the tap error signal, n for the tap additive noise, N for the tap count, and \(\theta^{(n)} = d_{opt} - d^{(d)}\) for the step size. If the definition of “tap error vector” is (n) = dopt

$$ \theta^{(n + 1)} = \theta^{(n)} - \alpha_{c} f(_{n} + v_{n} ;B^{(n)} )b^{(n)} $$
(6)

and

$$ e_{n} = b^{(n)T} \theta^{(n)} $$
(7)

The name of the proposed method suggests an adaptive control of the threshold \(B^{(n)}\). The long-term root mean square value \(e_{n} + v_{n}\) is used to determine how proportional BB should be.

Where M is a coefficient ranging from around 1–2, and \((\sigma_{L}^{(n)} )^{2}\) is the

$$ B^{(n)} = M\sigma_{L}^{(n)} $$
(8)

and

$$ (\sigma_{L}^{(n)} )^{2} = (1 - \beta_{L} )(e_{n} + v_{n} )^{2} + \beta_{L} (\sigma_{L}^{(n - 1)} )^{2} $$
(9)

The computation uses the long-term average of \((e_{n} + v_{n} )^{2}\) and include \(\beta_{L}\) s the leak factor inside the leaky accumulator to compute (9).

The filter convergence follows the following trend when the adaptive threshold constrained algorithm (ATCA) is used for adaptive filtering. The advanced telecommunications computing architecture (ATCA) is a standardized framework for signal processing and adaptive filtering because of its robustness and scalability. Its integrated management tools improve dependability by monitoring in real-time and optimizing parameters. ATCA prioritizes reliability with redundant components and fault tolerance, making it a preferred choice in these fields. Its adaptability enables parameter optimization for efficient adaptive filtering and signal processing. The error signal, along with \((\sigma_{L}^{(n)} )^{2}\) \(B^{(n)}\) and, tends to be significant during the early stages. Because the leak factor is so close to unity, BB remains considerable even as the filter gradually converges and the erroneous signal’s strength decreases. When the threshold parameter B surpasses the total size of the error and noise signal, its behavior resembles that of the least mean square algorithm (LMSA). In machine learning and signal processing, one common adaptive filtering method is the least mean square algorithm (LMSA). The least mean square algorithm (LMSA) is a widely used adaptive filter algorithm for minimizing mean squared error in real-time applications such as noise reduction and system identification. It iteratively adjusts filter weights based on the instantaneous error and input data using stochastic gradient descent. The update rule, convergence influenced by the learning rate, and applications in adaptive filtering and machine learning are important considerations. Its adaptability is increased by regularization and variations like Normalized LMSA. LMSA is a computationally efficient algorithm with adaptive step size options for dynamic adjustments. Its adaptability and simplicity make it popular for real-time tasks. Its capacity to iteratively modify filter coefficients, minimizing mean square error in changing system parameters, makes it popular for real-time applications.

The error signal becomes negligible, possibly even smaller than the power of the background noise, after convergence and achieving the ultimate mean squared value of en + n. As a result, in the steady state, the noise power is a few dB lower than the threshold \(B^{(n)}\).

As a result, the lower threshold effectively “clips” the peak of the impulse and preserves the majority of the residual error even if a high-power stimulation is applied to the system after convergence Fig. 3.

4 Experiments, Results, and Discussion

This section presents our proposed method’s extensive performance summary and experimental data.

4.1 Experimental Setup

A federated learning system based on evolutionary computation has been created to put the findings into practice for host intrusion detection in real-time traffic analysis. To detect intrusions on individual hosts within a network, the system analyzes network traffic in real-time. The practical application of these findings demonstrates the system’s ability to address real-world security challenges. Empowers real-time analysis of network traffic for swift responses to potential intrusions, bolstering overall security. Safeguards data privacy by localizing sensitive information on individual hosts, minimizing the risk of exposure. Emphasizes host-level detection of anomalies and potential security threats. The Raspberry Pi 3 Model B + is used for testing the proposed FL-HIDS. This Raspberry Pi model includes 1 GB of LPDDR2 SDRAM, a potent 1.4 GHz 64-bit quad-core processor, and compatibility with Bluetooth 4.2 as well as 2.4 GHz and 5 GHz IEEE 802.11b/g/n/ac wireless LAN. It also provides several other contemporary features crucial for the testing procedure. It is an edge computing tool and may be used with numerous ML models. The Raspbian operating system and many Linux variants power this gadget. Python, a high-level, multi-paradigm programming language, coupled with a few Python modules and TensorFlow (an open-source software library), is used to generate the ML source code. Two alternatives are provided for writing Python code: SSH and Mu editor. The development of an edge tensor processing unit for embedded and mobile devices to speed up TensorFlow calculations. The performance on edge devices is achieved by offloading machine learning inference tasks to specialized hardware designed for tensor operations in neural networks. This dedicated hardware, proficient in matrix multiplications, reduces the main processor workload, leading to faster TensorFlow computations at the edge. The system is optimized for effective performance and security using a 64-bit Ubuntu Linux operating system, an Intel(R) Core (TM) i5-8250U processor, 8.00 GB (7.89 GB useable), and the designated FL-HIDS.

4.2 Evaluation Metrics

A few measuring criteria are used to assess the suggested edge-based IDS. The performance of the model is evaluated using the subsequent measures. Precision, false alarm rate (FAR), recall, accuracy, attack detection rate (ADR), and F-score. TP, TN, FP, and FN are the four different measuring factors that must be included in the computation. In Table 2, each aspect is fully explained. Table 3 shows the evaluation metric’s complete description and computation equation.

Table 2 Measuring variables
Full size table
Table 3 Performance indicators
Full size table

4.2.1 Precision Analysis

In Fig. 4 and Table 4, the precision of the FL-HIDS method is compared with that of existing techniques. The graph demonstrates how the DL approach has increased efficiency with precision. For instance, the XGBoost, ADA, QDA, and LR models’ respective precision values for 100 data are 76.12%, 87.67%, 81.11%, and 91.45%, respectively, as opposed to the FL-HIDS model’s precision with 94.55%. However, the FL-HIDS model has performed best with various data. Similarly, under 500 data, the FL-HIDS has a precision of 96.12%, while the corresponding precision values for XGBoost, ADA, QDA, and LR are 80.45%, 90.56%, 86.45%, and 93.11%.

Fig. 4
figure 4

Precision analysis for FL-HIDS method

Full size image
Table 4 Precision Analysis for FL-HIDS method
Full size table

4.2.2 Recall Analysis

In Fig. 5 and Table 5, the recall of the FL-HIDS methodology is compared with that of other methods. The graph demonstrates how the DL approach has increased efficiency with recall. For instance, the XGBoost, ADA, QDA, and LR models’ respective recall values for 100 data are 67.12%, 72.44%, 79.34%, and 85.12%, respectively, as opposed to the FL-HIDS model’s recall value of 92.18%. However, the FL-HIDS method has been shown to perform best with various data. Similar to this, under 500 data, the FL-HIDS has a recall value of 97.45%, while the corresponding recall values for XGBoost, ADA, QDA, and LR are 71.43%, 76.66%, 83.44%, and 89.23%.

Fig. 5
figure 5

Recall analysis for FL-HIDS method

Full size image
Table 5 Recall analysis for FL-HIDS method
Full size table

4.2.3 F-Score Analysis

In Fig. 6 and Table 6, the f-score of the FL-HIDS method is compared with that of other methods. The graph demonstrates how the DL approach has increased efficiency with f-score. For instance, the XGBoost, ADA, QDA, and LR models’ respective f-score values for 100 data are 80.13%, 88.45%, 84.87%, and 90.67%, respectively, as opposed to the FL-HIDS model’s f-score value of 94.19%. However, the FL-HIDS method has been shown to perform best with various data. Like this, under 500 data, FL-HIDS has an f-score value of 97.12%, while the corresponding f-score values for XGBoost, ADA, QDA, and LR are 83.76%, 90.13%, 87.12%, and 93.87%.

Fig. 6
figure 6

F-score analysis for FL-HIDS method

Full size image
Table 6 F-score analysis for FL-HIDS method
Full size table

4.2.4 Accuracy Analysis

In Fig. 7 and Table 7, the accuracy of the FL-HIDS method is compared with that of other methods. The graph demonstrates how the DL approach has increased efficiency with accuracy. For instance, the XGBoost, ADA, QDA, and LR models’ respective accuracy values for 100 data are 82.67%, 88.12%, 85.19%, and 91.43%, respectively, as opposed to the FL-HIDS model’s accuracy of 96.98%. However, the FL-HIDS method has been shown to perform best with various data. Similar to this, under 500 data, the FL-HIDS has an accuracy of 99.90%, while the corresponding accuracy values for XGBoost, ADA, QDA, and LR are 84.45%, 90.34%, 87.98%, and 95.19%.

Fig. 7
figure 7

Accuracy analysis for FL-HIDS method

Full size image
Table 7 Accuracy analysis for FL-HIDS method
Full size table

4.2.5 Attack Detection Rate Analysis

In Fig. 8 and Table 8, the attack detection rate of the FL-HIDS method is compared with that of other methods. The graph shows how the DL approach has increased efficiency with attack detection rate. For instance, the XGBoost, ADA, QDA, and LR models’ respective attack detection rate values for 100 data are 74.12%, 82.78%, 79.13%, and 87.34%, respectively, as opposed to the FL-HIDS model’s attack detection rate of 92.87%. However, the FL-HIDS method has been shown to perform best with various data. Similar to this, under 500 data, the FL-HIDS has an attack detection rate of 96.78%, while the corresponding attack detection rate values for XGBoost, ADA, QDA, and LR are 78.88%, 86.12%, 81.87%, and 91.45%.

Fig. 8
figure 8

Attack detection rate analysis for FL-HIDS method

Full size image
Table 8 Attack detection rate analysis for FL-HIDS method
Full size table

4.2.6 False Alarm Rate Analysis

Figure 9 and Table 9 present a comparative analysis of the FL-HIDS method’s False Alarm Rate with other current methodologies. The DL approach has produced better performance with a lower False Alarm Rate, as the figure illustrates. The False Alarm Rate for FL-HIDS, for instance, is 23.19% with 100 data, whereas the XGBoost, ADA, QDA, and LR models have somewhat higher False Alarm Rates of 43.98%, 39.45%, 34.19%, and 29.76%, respectively. On the other hand, for various data numbers and low False Alarm Rate values, the FL-HIDS model has demonstrated maximum performance. Comparably, under 500 data, FL-HIDS’s False Alarm Rate is 27.77%, but the corresponding values for the XGBoost, ADA, QDA, and LR models are 47.87%, 42.87%, 38.88%, and 33.87%.

Fig. 9
figure 9

False alarm rate analysis for FL-HIDS method

Full size image
Table 9 False alarm rate analysis for FL-HIDS method
Full size table

5 Conclusion

This study explored deep learning (DL) methods frequently used in host intrusion detection systems (HIDS) that look for unusual activity to identify malicious data. The efficacy of DL-based techniques is increased by using the federated learning architecture, which enables several users to train a global model while protecting the privacy of each user’s data. However, a comprehensive investigation of how well FL-based HIDSs protect against recognized privacy concerns using existing protections has yet to be conducted. To address this issue, we provide two privacy assessment metrics for FL-based HIDSs, including a privacy score that quantifies how similar the original and restored traffic attributes are. The real-time model was built using the CICIDS2017 dataset, which includes numerous recent attacks. An Adaptive Threshold-correlation Algorithm (ATCA) is also introduced to improve detection accuracy by dynamically altering threshold values based on traffic patterns and incursion behaviors. The FL-HIDS framework was created and tested using a realistic network dataset. Studies show that the suggested method outperforms the most cutting-edge intrusion detection systems regarding detection accuracy and scalability. The federated learning technique successfully harnesses the network device’s collective intelligence, enabling continuous learning and adaptation to change attack strategies. In terms of system performance, the suggested method beat five currently used techniques, including simultaneous XGBoost, Ada boost classifier (ADA), quadratic discriminant analysis (QDA), and logistic regression (LR). According to research findings, the FL-HIDS ideal has average precision rates of 96.12%, recall rates of 97.45%, f-scores of 97.12%, accuracy rates of 99.90%, attack detection rates of 96.78%, false alarm rates of 27.77%. Future projects will encompass a range of activities. For network security operations, analyzing datasets suited for replicating vertically partitioned data is essential since it aids in assessing how well-federated learning algorithms function. The authors want to investigate approaches that make it easier to map various feature sets into a single feature space because of the tight relationship between this objective and another potential research direction.