1.1 Relevance of the N-BaIoT Dataset

To evaluate the performance of FedAvg for IoT botnet detection, we used the N-BaIoT dataset[11], which is a real-world benchmark containing benign and malicious network traffic from multiple IoT devices infected with Mirai and Gafgyt botnets [12]. This dataset provides diverse attack scenarios, making it well-suited for assessing the robustness of FL-based anomaly detection. An appropriate citation of the dataset was included to ensure reproducibility.

1.2 Key Contributions

This study makes the following key contributions.

The remainder of this paper is structured as follows. Section 2 reviews related work on IoT botnet detection and Federated Learning. Section 3 describes the proposed methodology, including data pre-processing, model architecture, and FL implementation. Section 4 presents experimental results and performance analysis. Section 5 discusses key findings, limitations, and future research directions. Finally, Section 6 concludes this paper.

2.1 Federated Learning for IoT Security

IoT networks present unique security challenges because of their distributed nature, heterogeneous devices, and constrained computational resources [2, 1]. Traditional centralized machine-learning solutions for securing IoT networks require raw data aggregation, which increases the risk of privacy breaches and scalability limitations [4, 15, 16]. FL has emerged as an effective approach for mitigating these risks by enabling decentralized model training while maintaining data localized on edge devices [7, 6].

Konečný et al. [5] examined the advantages of FL in securing IoT ecosystems, highlighting its potential to preserve privacy while reducing communication overhead. Yang et al. [17] conducted an extensive survey on FL applications, emphasizing their role in securing IoT networks without exposing raw data. Additionally, techniques such as adaptive aggregation [18], differential privacy [19], and secure multiparty computation [20] have been integrated into FL frameworks to enhance security.

However, most existing studies have focused on general anomaly detection rather than evaluating FL specifically for botnet threats. For example, Li et al. [21] and Xu et al. [22] demonstrated FL's effectiveness of FL in anomaly detection but did not assess its applicability to botnet detection. This study fills this gap by evaluating FL specifically for botnet detection using real-world botnet traffic data.

2.2 Federated Learning for Anomaly Detection

Anomaly detection is essential for IoT security because it enables the identification of malicious activities such as botnet attacks, unauthorized access, and data exfiltration [23, 24]. Several deep-learning-based anomaly detection frameworks have been proposed, but they often rely on centralized data collection, which is impractical in privacy-sensitive environments [25, 4].

Xu et al. [22] applied FL to medical anomaly detection and demonstrated its ability to train robust models while preserving data privacy. Similarly, Mothukuri et al. [26] explored FL-based anomaly detection for edge devices, illustrating its potential to reduce computational load while maintaining accuracy. Despite these advances, FL for IoT botnet detection remains underexplored. Studies such as [26] evaluated FL for general security threats, but did not specifically consider the Mirai and Gafgyt botnets, which are among the most prevalent IoT botnet families. This study addresses this gap by evaluating botnet detection using the N-BaIoT dataset while also analyzing the performance under IID and Non-IID data distributions.

2.3 Botnet Detection Using Federated Learning

The detection of botnet attacks in IoT networks has been studied extensively, with existing solutions leveraging traditional machine learning, deep learning, and statistical methods [27, 28]. Conventional approaches rely on network traffic analysis using centralized models, which require significant computational resources and exposure to sensitive data [15, 24].

Meidan et al. [11] introduced the N-BaIoT dataset, demonstrating the feasibility of deep learning-based botnet detection. However, their study relied on a centralized training approach, which limits its real-world applicability. More recently, Popoola [29] proposed an FL-based botnet detection framework, showing that FL models can achieve accuracy levels comparable to those of centralized models while preserving privacy. Similarly, Xiong et al. [30] studied FL for botnet detection under non-IID conditions, highlighting the performance gap compared to IID scenarios.

Unlike traditional signature-based and anomaly based detection methods[31, 32], which require frequent updates and retraining, FL is a promising alternative. By continuously learning from distributed attack patterns without centralized data collection [6, 29], FL improves the adaptability to evolving botnet threats.

2.4 Limitations of Existing Work

Despite the progress in FL-based botnet detection, several challenges remain.

This study directly addresses these gaps by:

3.1 N-BaIoT Dataset

The N-BaIoT dataset[11] is a benchmark dataset containing the network traffic from IoT devices infected by Mirai and Gafgyt botnets. The dataset includes the following attack categories.

3.2 Federated Learning Implementation

The FL framework was implemented using PyTorch and Flower following the FedAvg algorithm [10].

 Federated Learning Workflow

Federated Learning (FL) enables distributed training across IoT devices without transmitting raw data to a central server, thereby preserving the privacy. The training process followed the iterative Federated Averaging (FedAvg) mechanism.

Figure 1 illustrates the FL workflow, which consists of the following steps.

• Global Model Initialization: The central server initializes a global model and sends it to all FL clients (IoT devices).

• Local Training: Each client trains the model using its local dataset without sharing raw data.

• Model Update Transmission: After training, each client sends its model updates (weights) to the central server.

• Model Aggregation (FedAvg): The server aggregates received model updates using the Federated Averaging (FedAvg) algorithm:

  $$w_{t+1}=\sum _{k=1}^K\frac{n_k}{N}{w}_k^t$$
  where:

  • $w_{t+1}$ is the global model weight at round $t+1$.

  • $K$ is the number of participating clients.

  • $n_k$ is the number of training samples at client $k$.

  • $N$ is the total number of training samples across all clients.

  • $w_k^t$ is the local model weight of client $k$ at round $t$.

• Global Model Update: The aggregated model is sent back to clients for the next training round.

• Convergence Check: The FL process continues until the model reaches a predefined accuracy threshold.

Figure 1 Federated learning workflow.

3.3 Model Architecture

The IoT botnet detection model is a deep neural network (DNN) designed for the binary classification of network traffic as either benign or botnet attacks. The architecture balances the computational efficiency and accuracy.

Figure 2 illustrates the proposed neural network architecture for botnet detection. The model consists of an input layer (115 features), three hidden layers with ReLU activation and batch normalization, and a softmax output layer for binary classification. Dropout layers (30% rate) are applied after each hidden layer to mitigate overfitting.

Figure 2 Neural network model architecture.

3.4 Experimental Setup

3.5 Communication Efficiency Optimization

The following techniques were implemented to reduce communication costs:

3.6 Comparison with Centralized Learning

FL is compared with a centralized approach, in which all IoT data are aggregated for training. The key metrics include the following.

4.1 FL vs Centralized Learning: Accuracy and Convergence

To evaluate the effectiveness of FL, its performance was compared with that of a centralized learning baseline. Figure 3 shows the accuracy trends over multiple training rounds.

Figure 3 Comparison of Accuracy vs. Training Rounds for FL (IID, Non-IID) and centralized learning.

Key Findings:

Statistical Significance: A paired t-test indicates a statistically significant difference () in convergence speed between FL (IID) and FL (Non-IID), confirming that Non-IID settings introduce learning delays.

4.2 Impact of Non-IID Data Distribution

FL models are affected by non-IID data distributions, where clients train on non-uniform datasets. Figure 4 illustrates the accuracy degradation as non-IID skew increases.

Figure 4 Impact of Non-IID data distribution on FL accuracy.

Quantifying Non-IID Skew:

Key Findings:

4.3 Performance Metrics for Botnet Detection

Table 1 presents the precision, recall, and F1-score results for the FL and centralized models.

4.4 Impact of Communication Overhead Reduction

Table 2 evaluates the impact of gradient sparsification and quantization on the communication overhead.

4.5 Computational Cost Analysis

Table 3 compares the training times and memory usage of the FL models.

5.1 Summary of Key Findings

This study demonstrates that FL can effectively detect botnet attacks while preserving data privacy. The key findings are as follows.

5.2 Challenges and Limitations

Despite its benefits, FL presents several challenges when applied to IoT-botnet detection.

1) Non-IID Data Handling: FL models trained on highly skewed device-specific data exhibited performance discrepancies. Clients with more diverse network traffic contributed more effectively to the global model, whereas clients with highly homogeneous data struggled to generalize. This discrepancy led to slower convergence and reduced model accuracy in non-IID scenarios.

2) Communication Efficiency vs. Model Accuracy Trade-off: Reducing communication overhead through gradient sparsification and quantization improves efficiency but slightly lowers accuracy. Optimizing this balance remains an open research question.

3) Computational Constraints: Many IoT devices have limited processing power and memory, which may restrict the feasibility of deploying complex deep learning models in real-world settings.

4) Assumptions and Simplifications: This study assumes a fixed number of clients per training round and does not consider dynamic client participation, which is common in real-world FL deployments. Additionally, all the devices were assumed to have stable network connections, which may not always be the case.

5.3 Potential Improvements and Future Work

To address these challenges, several future research avenues can be explored, ordered according to their impact and feasibility.

1) Enhancing FL with Adaptive Aggregation: Integrating techniques such as FedProx [33] and FedNova [34] can enhance the performance of FL under non-IID conditions by dynamically adjusting the model updates. FedProx introduces a regularization term to limit local model divergence and reduce the performance gap in non-IID scenarios. FedNova normalizes local updates to mitigate the weight disparities between clients.

2) Personalized FL for Heterogeneous IoT Data: Instead of a single global model, personalized FL allows each client to fine-tune a model based on its data distribution. Meta-learning techniques (e.g., Model-Agnostic Meta-Learning, MAML) could be explored to enable client-specific adaptations.

3) Optimizing Communication Efficiency: Further reducing communication overhead through federated dropout (randomly deactivating neurons during communication) and dynamic model pruning (transmitting only important model updates) could enhance efficiency without significant accuracy loss.

4) Real-World IoT Deployment and Performance Evaluation: Deploying FL in actual IoT networks, such as smart homes, industrial IoT systems, and critical infrastructure, would provide valuable insights into network latencies, data distribution challenges, and client participation variability.

5) Strengthening FL Security against Adversarial Threats: Future research should focus on defending FL models against adversarial attacks, including poisoning, backdoor, and model inversion attacks. Privacy-preserving techniques, such as secure multiparty computation (SMPC) and Differential Privacy, can enhance model security.

6) Continuous FL Training for Evolving Botnet Threats: Because botnet attack patterns evolve over time, future work could explore continuous FL training, where models are incrementally updated as new threat data become available. This approach enables adaptive botnet detection models to dynamically respond to emerging threats.