SA-FLIDS: secure and authenticated federated learning-based intelligent network intrusion detection system for smart healthcare

Blockchain

Is a distributed ledger technology for data transmission and storage that records the history of all transactions. Blockchain has been involved due to its decentralization, immutability, and persistence properties in the distributed peer-to-peer (P2P) network. It utilizes asymmetric cryptography to ensure transactions are done safely (Saidi et al., 2022). Blockchain is based on several elements used to create a secure, transparent, and decentralized system. Each element plays a crucial role in maintaining the integrity and functionality of the blockchain (Mboussam Emati & Mboussam, 2023), including block, transaction, consensus mechanism, smart contract, mining, immutable ledger, cryptographic keys, and hash function.

Self-sovereign identity

is a decentralized approach and a new model for digital identity. Self-sovereign identity (SSI) aims to empower individuals to possess and control digital proof of their credentials. Thus, it helps to prove who we are by establishing trusted relationships to access information (Saidi et al., 2022). SSI is based on two main standards: Decentralized identifier (DID) and verifiable credential (VC):

• Decentralized identifier is a new type of identifier, defined by the W3C (Emati, Mboussam & Tchendji, 2023). DIDs are designed to enable Self-Sovereign Identity on the internet, providing a way for individuals to have control over their own digital identity without the need for a central authority. Thus, users can selectively disclose only the necessary information for a particular transaction or interaction (Saidi et al., 2022).

• Verifiable credential is a standard method for digitally expressing credentials in a cryptographically secure way. It can include metadata, claims, and proofs used to verify a credential (Thomas, Ramaguru & Sethumadhavan, 2022). Credentials are created and signed by the issuer using his private key and then issued to the holder, enabling the verifier to confirm the VC. A holder keeps and shares the received credentials with a verifier. The verifier accepts and approves these credentials (Figueroa-Lorenzo, Benito & Arrizabalaga, 2021). based on the public key associated with a DID.

Trimmed-mean aggregation method

The Trimmed-Mean method enhances the accuracy and reliability of the global model in FL by mitigating the influence of outliers and malicious participants. It removes a certain percentage of extreme values from local models and calculates a weighted average of the remaining models. The Trimmed-Mean formula is:

(1) Trimmed−Mean=∑i(wi⋅mi)∑iwiwhere:

wi represents the weight of the i-th local model.

mi represents the i-th local model in the trimmed set Mt.

Experimental setup for federate learning

Table 1 presents the parameters used in federated deep learning. In our study, we conducted experiments deploying our model with client sets denoted as K, where K = 10. We employed the Independent and Identically Distributed (IID) approach, ensuring that the data distribution across the dataset matches the distribution of data for each client. Furthermore, To prevent overfitting, the following techniques were used:

• Stratified K-fold cross-validation: Set k = 5 to split the data into five subsets, evaluating the model’s performance.

• L2 regularization: Applied with a factor of 0.01 to the dense layers, adding a penalty to the loss function based on the weights’ magnitude, simplifying the model.

• Early stopping: With the patience of three, training stops if the validation loss does not improve, monitoring the validation set’s performance.

Experimental setup for blockchain-based SSI

The simulation is carried out on a Lenovo ThinkPad P51 -Core i7 2.9 GHz–SSD 1 To–32 Go, computer running Ubuntu 18.04 LTS. The simulation environment is based on Hyperledger Indy (Banerjee et al., 2022), a framework dedicated to self-sovereign identity management. It offers an abstraction that enables DIDs and VCs to be created, verified, and revoked. It embeds Hperledger ursa, a module that provides all the primitives for cryptographic operations. We also use Hyperledger Aries (Manoj, Makkithaya & Narendra, 2022). This is a library for creating agents that can manage the cryptographic wallet of each player. It also offers interfaces for creating functionalities that realize the behavior of the wallet owner.

The simulation begins with the initialization phase, during which 10 nodes plus the server are each initialized in a Docker container as shown in Fig. 6. The server is the trusted authority that manages the blockchain. It creates the genesis block and initializes the chain. Each node is registered on the chain and receives a DID and an authentication scheme. Of course, each DID is accompanied by its key pair, which is stored in each node’s wallet, and the public keys are known to everyone. Next comes the authentication phase, when each node sends its VC to the server, which verifies it and receives the server’s VC (mutual authentication). The principle is to request the blockchain’s response to the VC presented to it. Once each node’s authenticity has been verified. The nodes can exchange data. Remember that confirmation of a VC leads to validation of a token, which has a lifetime (that of the session). Each node in possession of its token can then use it for future exchanges. After authentication, the server sends the chunks of the dataset to the various nodes using its private key as demonstrated in Fig. 3. The nodes receive them and can proceed with processing. The different scores are exchanged using the same process. The communication protocol in this phase is gRPC. However, the solution would work with any other protocol, with security being guaranteed by the authentication token.

Datasets description

Datasets play an essential role in both training and assessing IDSs within IoT networks. The choice of suitable datasets tailored to particular tasks holds significant importance, particularly in evaluating the efficacy of FL approaches for IoT networks. In our experiment, we incorporated two recent datasets specifically designed to mimic real-world conditions for IDSs: CICIoT2023 (Neto et al., 2023), made available in 2023, and the Edge-IIoTset dataset, released in 2022.

1. CICIoT2023 dataset: A Neto et al. (2023) novel and extensive IoT attack dataset to foster the development of security analytics applications in real IoT operations. To accomplish this, 33 attacks are executed in an IoT topology composed of 105 devices, and all attacks are executed by malicious IoT devices targeting other IoT devices. We analyzed a dataset containing 47 features (not including label and sublabel) based on 2,366,956 samples extracted from the first 10 CSV files provided by the Canadian Institute.

2. Edge-IIoTset dataset: It is tailored specifically for IIoT and IoT applications (Ferrag et al., 2022), providing an authentic test environment closely resembling real-world IoT/IIoT settings. Within this environment, we conducted simulations of genuine cyberattacks to collect datasets comprising both legitimate and malicious network traffic. This dataset includes data generated by various IoT devices, spanning from heart rate sensors to flame sensors, temperature, and humidity sensors. The testbed is structured into seven interconnected layers. We utilized the Selected dataset for ML and DL/DNN-EdgeIIoT-dataset CSV file (Banerjee et al., 2022), which contains 61 features and 2,219,201 samples, encompassing both normal traffic and 14 distinct attacks in the IoT and IIoT environment.

Preprocessing

The datasets undergo several preprocessing steps to ensure their suitability for analysis as demonstrated in Fig. 7. After cleaning the data we first, address imbalanced data by implementing SMOTE (Synthetic Minority Over-sampling Technique) and under-sampling techniques to enhance predictive performance, particularly for minority classes. Secondly, data transformation is conducted using the StandardScaler for standardization, adjusting data to have a mean of 0 and a standard deviation of one. Additionally, feature importance analysis is performed using insights from random forest and XGBoost experiments. Finally, the processed dataset is split into an 80% training set and a 20% testing set, ensuring no duplication between the two, contributing to refining the dataset for subsequent analysis and modeling tasks. In the case of the CICIoT dataset, we opt to eliminate the Brute and Web attack labels due to their limited number of samples, which could potentially skew the analysis and compromise the reliability of the results. The detailed features selected and the attacks used are outlined in Table 2 provided below.

Metrics used for federated learning evaluation

When conducting intrusion detection using federated deep learning performance analysis, the most common metrics used are:

• True negatives (TN): Benign network activity correctly classified as normal.

• True positives (TP): Malicious network activity correctly identified as an attack.

• False positives (FP): Benign network activity incorrectly classified as malicious.

• False negatives (FN): Malicious network activity is incorrectly classified as normal.

Moreover, we have used a variety of measures to evaluate our proposed model, including precision, recall, precision, F-score, and accuracy, to conduct a systematic comparative analysis with other relevant approaches as demonstrated in Table 3.

Metrics used for blockchain-based SSI evaluation

To compute the metrics outlined below, we utilize the following formulas:

1. Startup duration (SD): The duration for the system to initiate.

2. Connect duration (CD): The time required for the system to establish connections between nodes and the Fog server.

3. Publish duration (PD): The duration for the system to publish schema credentials and related settings.

4. Issuing credential duration (ICD): The time taken for the system to issue credentials.

5. Completed credential exchanges duration (CCED): The total time needed for all credential exchanges to conclude.

   Metrics such as SD, CD, PD, and ICD are used to assess the Initialization phase, while CCED is used to assess the Mutual Authentication phase.

6. Average time per credential duration (ATCD): The average time taken to issue a single credential.

   (2) ATCD=ICDNcredentials

   where:

   ICD is the issuing credential duration.

   N_credentials is the total number of credentials issued.

7. Average time per transaction duration (ATTD): The average time taken per transaction.

   (3) ATTD=∑i=1NtransactionsTiNtransactions

   where:

   N_transactions is the total number of transactions.

   T_i is the duration of each transaction i.

Binary classification

In this subsection, we present the evaluation results for binary classification scenarios using FL. In addition, we provide an evaluation of blockchain-based SSI.

1. Federated learning evaluation results. We employed 150,000 samples for both benign and attack instances in both datasets, ensuring a balanced dataset for a comprehensive and meaningful comparison. Remarkably, our model demonstrates impeccable performance, achieving perfect scores of 100% across all metrics for the Edge-IIoT dataset. In contrast, the results for the CICIoT2023 dataset remain highly promising, with an accuracy of 99.09%, indicating a low error rate in classifying both benign and malicious traffic. Furthermore, achieving a perfect precision of 100%, along with a recall of 98% and an F1-score of 99%, underscores the robust overall performance of the model, as shown in Fig. 8.

The classification performance of our model is depicted through the confusion matrix presented in Fig. 9, providing a concise summary of the model’s accurate and erroneous predictions The primary goal is to minimize both false positive and false negative rates, ensuring precise classification outcomes. Our proposed model effectively achieves this objective, exhibiting false positive and negative rates of 0% in the Edge-IIoTset dataset. For the CICIoT2023 dataset, we observe a negligible false negative rate of 0.0017%, alongside false positive rates of 0. 96%, which confirms the accuracy and efficiency of the model in mitigating classification errors.

Furthermore, the receiver operating characteristic (ROC) curve and the area under the curve (AUC) depicted in Fig. 10 offer a visual representation of our model’s ability to distinguish between classes to highlight the model’s effectiveness, we achieve an AUC of 99.1% for CICIIoT2023 notably 100% for the EdgeIIoTset dataset.

The training and validation loss curves for both the EdgeIIoTset and CICIoT2023 datasets demonstrate the promising performance of our federated learning model in binary classification. As illustrated in Fig. 11, the models exhibit rapid convergence within the initial epochs, followed by stable performance. The close alignment of training and validation loss curves, particularly in later epochs, indicates good generalization without significant overfitting. Both models achieve remarkably low final loss values (¡ 0.02) for training and validation sets, suggesting high predictive accuracy. The EdgeIIoTset model shows a slightly lower final loss, while the CICIoT2023 model displays smoother convergence between training and validation losses. These results collectively suggest that our approach effectively captures the underlying patterns in both datasets, promising strong performance on unseen data in real-world applications.

2. SSI-based DID evaluation results. As illustrated in Fig. 12 below both datasets exhibit similar startup durations (SD), with EdgeIIoTset demonstrating a marginally quicker performance by 0.01 s. Furthermore, the EdgeIIoTset dataset shows a shorter connect duration (CD) of 0.02 s compared to the CICIoT2023 dataset. Both datasets share the same publish duration (PD) of 9.15 s. However, in terms of issuing credential duration (ICD), the EdgeIIoTset dataset outperforms the CICIoT2023 dataset by 0.87 s. Regarding completed credential exchange duration (CCED), the EdgeIIoTset dataset exhibits a reduction of 8.9 s compared to the CICIoT2023 dataset. the CCED should normally be higher because this is the mutual authentication phase since the FS must authenticate all the nodes, and each node must authenticate the FS. Additionally, the average time per credential duration (ATCD) is shorter for the ‘EdgeIIoTset’ dataset in comparison to the ‘CICIoT2023’ dataset. Finally, the average time per transaction duration (ATTD) is marginally higher for the ‘EdgeIIoTset’ dataset when contrasted with the ‘CICIoT2023’ dataset.

Multiclass classification

In this subsection, we present the evaluation results for both the CICIoT2023 and EdgeIIoTset datasets in multiclass classification scenarios using FL. Additionally, we provide an evaluation of SSI-based DID.

1. Federated learning evaluation results. The evaluation of both the CICIoT2023 and Edge-IIoTset datasets reveals strong performance across diverse classes, as demonstrated in Table 4. Within the CICIoT2023 dataset, each class achieves good performance in most metrics. Notably, the DDoS, DoS, and Mirai attack classes in the CICIoT2023 dataset exhibit great classification capabilities, demonstrating perfect performance across all metrics with complete precision, recall, and F1-scores of 100%. Furthermore, the Spoofing, Benign, and Information Gathering classes show good precision with 87%, 83%, and 81%, respectively. However, their recall and F1 scores vary. The Benign class achieves a high recall of 85% and an F1-score of 84%. The Information Gathering class presents a relatively good recall of 84% and a corresponding F1-score of 83%. The Spoofing class achieves a moderate recall at 80% and an F1-score of 84%.

Transitioning to the Edge-IIoTset dataset Man in the Middle, and Normal classes also exhibit high performance, achieving perfect precision, recall, and F1-scores of 100%. followed by the DoS/DDoS class with high precision 91% and very high recall 99%, resulting in a strong F1-score of 92%. Furthermore, other classes show more variability. The Information Gathering class has an excellent precision, recall, and f1-score with 88%. The Injection class shows a precision of 77% and a high recall of 91%, leading to an F1-score of 84%. The Malware Attacks class, despite its high precision of 97%, suffers from a low recall of 50%, resulting in a lower F1-score of 74%.

The confusion matrix depicted in Fig. 13 provides valuable insights into the prediction frequency for each class compared to their actual occurrences. Within the CICIoT dataset, remarkable performance was observed for classes such as DoS, DDoS, and Mirai, with the model accurately predicting all samples, achieving a flawless prediction rate of 100%. Additionally, the classification of Reconnaissance exhibited high accuracy, with 84% of samples correctly classified followed by Benign traffic with 85%. However, in subsequent classes such as Spoofing, accuracy decreased, with only 80% respectively, of the samples accurately classified. Likewise, within the Edge-IIoTset dataset, classes like Man in the Middle, and Normal traffic demonstrated robust performance, as the model accurately predicts all samples, resulting in a 100% prediction rate. The classification of the DoS/DDoS, Injection, and Information gathering class followed suit, with 94%, 91%, and 88% respectively of samples correctly classified. Nevertheless, as we delve into subsequent classes such as Malware, declined, with only 50%, respectively, of samples accurately classified.

The ROC curves and AUC values for both datasets demonstrate excellent detection capabilities across various attack types as presented in Fig. 14. In the EdgeIIoTset dataset, most attacks show very good performance with AUC values above 0.90, with “Man in the middle” and “Normal” attacks achieving perfect detection (AUC 1.00). While “Malware” has the lowest performance (AUC 0.75), other attacks like “DoS/DDoS”, “Information gathering”, and “Injection” exhibit strong performance (AUC 0.96-0.93-.94). The second CICIoT2023 dataset similarly shows outstanding results, with DDoS, DoS, and Mirai attacks reaching perfect detection (AUC 1.00). “Benign” and “Recon” categories perform very well (AUC 0.90-0.91), and even the lowest performing “Spoofing” category maintains good detection ability (AUC 0.88). Notably, all attack types in both datasets significantly outperform the random guessing baseline, indicating robust and effective detection across the board.

Figure 15 illustrates the training and validation loss curves for both the EdgeIIoTset and CICIoT2023 datasets demonstrating the promising performance of our federated learning model in multi-class classification. For the EdgeIIoTset, we observe a rapid initial decrease in both training and validation loss, followed by a gradual convergence. The CICIoT2023 dataset shows a more gradual, consistent decrease in both losses across epochs. Importantly, neither dataset exhibits signs of overfitting, as the validation loss continues to decrease alongside the training loss, with only minimal divergence in later epochs. The EdgeIIoTset model achieves slightly lower final loss values (around 0.22) compared to the CICIoT2023 model (about 0.24), suggesting robust performance across different IoT datasets. The close alignment between training and validation losses, particularly in later epochs, indicates good generalization capabilities of our federated learning approach. These results suggest that our model effectively learns from local data without compromising privacy while maintaining strong predictive performance across diverse IoT classification tasks.

2. SSI-based DID evaluation results. As illustrated in Fig. 16 both datasets exhibit similar startup durations (SD), with EdgeIIoTset demonstrating a slight advantage of 0.08 s. The CICIoT2023 dataset shows a shorter connect duration (CD) by 0.01 s compared to EdgeIIoTset. Additionally, both datasets share the same publish duration (PD) of 9.15 s. However, the CICIoT2023 dataset boasts a shorter issuing credential duration (ICD) of 0.29 s compared to EdgeIIoTset. Regarding the completion of the credential exchange duration (CCED), the CICIoT2023 dataset surpasses EdgeIIoTset by 5 s. The CCED should normally be higher because this is the mutual authentication phase since the FS must authenticate all the nodes, and each node must authenticate the FS. Furthermore, the CICIoT2023 dataset demonstrates a shorter average time per credential duration (ATCD) by 0.36 s compared to EdgeIIoTset. Lastly, the CICIoT2023 dataset shows a shorter average time per transaction duration (ATTD) by 3 s compared to the EdgeIIoTset.

In Table 5 we provide a comparison between the performance of our work with other FL-based state-of-the-art IDS. The proposed SA-FLIDS demonstrates superior performance compared to existing state-of-the-art FL-based IDS approaches. It achieves the highest accuracy of 100% for binary classification on the EdgeIIoTSet dataset with a standard deviation σ = 0.00%, outperforming previous methods. For multiclass classification, SA-FLIDS attains 93.48% accuracy and σ = 0.12% on EdgeIIoTSet and 92% and σ = 0.47% on CICIoT2023, surpassing earlier works. The model’s consistently high performance across different datasets containing emerging new cyber attacks on IoT networks and classification tasks underscores its robustness and effectiveness in intrusion detection for IoT environments.