Secure Data Sharing in Federated Learning through Blockchain-Based Aggregation

: In this paper, we explore the realm of federated learning (FL), a distributed machine learning (ML) paradigm, and propose a novel approach that leverages the robustness of blockchain technology. FL, a concept introduced by Google in 2016, allows multiple entities to collaboratively train an ML model without the need to expose their raw data. However, it faces several challenges, such as privacy concerns and malicious attacks (e.g., data poisoning attacks). Our paper examines the existing EIFFeL framework, a protocol for decentralized real-time messaging in continuous integration and delivery pipelines, and introduces an enhanced scheme that leverages the trustworthy nature of blockchain technology. Our scheme eliminates the need for a central server and any other third party, such as a public bulletin board, thereby mitigating the risks associated with the compromise of such third parties.


Introduction
With the advancement of big data and artificial intelligence (AI) technologies, the challenge of securely and reliably training machine learning (ML) models on distributed and heterogeneous data sources without compromising privacy and integrity has become increasingly prominent.The term federated learning (FL) was introduced by Google in 2016, at a time when the use and misuse of personal data were gaining global attention.The Cambridge Analytica scandal awakened Facebook users and those of similar platforms to the dangers of sharing personal information online [1].It also sparked a wider debate on the pervasive tracking of people on the internet, often without their consent.In response, many countries and regions have passed or proposed data privacy laws, such as the General Data Protection Regulation (GDPR) in Europe [2].
FL is a distributed ML paradigm that enables multiple entities to collaboratively train a model from their local data, without exposing their raw data to each other or a central server [3].This approach stands in contrast to traditional centralized ML techniques, where local datasets are merged into one location.Therefore, FL has the potential to address the prevalent limitations and challenges in the traditional approach, particularly the critical issues of data privacy.The main application of FL is to train models on data that are sensitive, distributed, or heterogeneous, such as personal data on mobile phones or data from different organizations.For example, FL can be used to improve spam filters and recommendation tools without accessing users' emails or preferences.It can also be used to leverage data from sensors and smart devices for various tasks, such as anomaly detection, predictive maintenance, and optimization.
Despite the promises of FL, two main risks persist in the distributed and decentralized nature of the learning process, namely the privacy challenge and model quality.Privacy concerns arise from adversaries attempting to identify sensitive patterns in local updates, potentially compromising the global model [4][5][6].Lyu et al. [7] provide a general discussion about privacy and robustness attacks against FL and provide some direction for mitigating them.The German BSI published a detailed report on specific privacy risks, such as membership inference attacks and model inversion attacks [8].Simultaneously, in terms of model quality, integrity challenges stem from participants acting maliciously by injecting poisoned or malformed inputs during the FL process, where even a single malicious input can have a significant impact [9,10].Apart from these risks, there are also other related concerns for FL in practice, and a comprehensive survey on them is provided in [11].To mitigate these risks, various solutions have been proposed.For example, with secure aggregation, instead of sharing the plaintext local update, only masked updates are transmitted, so that the server can aggregate the global model correctly without the knowledge of individual updates [12,13].In addition, differential privacy is also widely used, adding noise to the exchanged values to provide additional robustness [14,15].In the literature, one notable contribution to addressing these challenges is the EIFFeL framework [16].By leveraging a public bulletin board, EIFFeL aims to preserve the privacy of input data from different clients as well as guarantee the integrity of the inputs from these clients (e.g., the value of inputs should be in proper ranges).

Contribution and Organization
As privacy and integrity are two key challenges hindering the widespread adoption of FL, this paper is dedicated to investigating a rigorous solution to address these challenges.
Referring to [11], FL can be categorized into two categories: cross-silo FL and cross-device FL.In this paper, we focus on the cross-silo setting, wherein model training occurs on siloed data belonging to several clients from different organizations.These clients can communicate with third-party servers in a synchronous manner.
Our first contribution involves investigating the architecture and workflow of the EIFFeL framework along with analyzing its security guarantees.To this end, we identify specific areas where the EIFFeL framework exhibits potential risks, such as the compromise of the central server and/or the public bulletin board.We also highlight other efficiency concerns.
Our second contribution involves adapting the EIFFeL framework to leverage the security guarantees of blockchain platforms.To this end, we adapt the operations in the EIFFeL framework so that the functions of both the central server and the public bulletin board can be replaced by a blockchain platform.To address the challenge posed by blockchain's limitation to only perform deterministic operations and the requirement that no direct communication among clients is necessary, we integrate the Burmester-Desmedt group key exchange protocol into our solution, to make sure that the clients can jointly compute the final parameter update without breaching its confidentiality.
Our last contribution involves implementing our solution and shedding some light on the computational complexities.Note that the proposed solution is blockchain platform agnostic, so we omit the implementation details related to the blockchain part.This is earmarked for future work.
It is worth noting that privacy protection for FL is a very challenging task.Even with secure aggregation, some privacy risks may still remain [11].To address this, differential privacy, particularly local differential privacy, can be applied.Since such measures are used to secure aggregation and can be applied independently, we leave this out in this paper.
The paper consists of the following sections: we present the background and motivation of FL and blockchain in Section 2. In Section 3, we review the design and analysis of the EIFFeL framework.Following this, an enhanced scheme is proposed; its security analysis, along with a demonstration of performance, is presented in Section 4. Finally, we summarize our work in Section 5.

Preliminary 2.1. Federated Learning Approach
FL represents a decentralized paradigm that is reshaping traditional ML methods.In this collaborative approach, multiple clients contribute to training a shared model without divulging raw data.Assuming a central server and n clients C i (1 ≤ i ≤ n), a high-level overview of the FL workflow is summarized in Figure 1; we refer to the relevant handbook for more details, such as [3] Model update: Each client, C i , resets its local model with the parameters, U .As we mentioned in Section 1, while FL enhances privacy by not sharing raw data with the central server or other devices, a crucial concern arises: the trustworthiness of the central server.Two major challenges need to be considered: privacy challenge and integrity challenge, as explained in Section 3.

Blockchain Overview
A blockchain is a chain of blocks; each block contains a list of transactions.These blocks are linked and secured through cryptographic hashes, forming a continuous, unalterable chain.The decentralized nature of a blockchain means that no single entity has control over the entire network, mitigating the risk of a central point of failure.Transactions in a blockchain are grouped into blocks, and each block includes a reference to the previous block, creating a chronological chain of events.Miners, i.e., participants in the network with computational power, play a crucial role in validating and adding transactions to the blockchain.This sequential structure ensures the integrity of the data, as altering information in one block would require changing all subsequent blocks, which is an impractical and computationally infeasible task.For a more comprehensive review, we refer to comprehensive references, such as [17][18][19].Irrespective of the variations in blockchain structures, the following properties are anticipated: -Democracy and decentralized control: In systems utilizing proof of work (PoW) as the consensus mechanism in permissionless scenarios, everyone has the potential to act as a miner, possessing equal privileges to generate and approve blocks for the blockchain.
Although variations may exist in different cases, the overarching principle remains: blockchain technology eliminates the need for a singular, fully trusted entity, thereby averting the vulnerability of a single point of failure.-Integrity and immutability: In the absence of an attacker or a coalition of attackers dominating the consensus process, such as when more than 51% of the computing power in the Bitcoin blockchain is handled by semi-honest miners, it becomes infeasible to modify agreed-upon blocks in the consensus.
-Consistency: Despite potential attacks from robust adversaries, the chain upholds a singular and consistent perspective, as outlined by the aforementioned assumptions.However, it is essential to note that deviations from predefined rules by nodes can lead to the generation of forks, resulting in different perspectives among participants, as observed in scenarios like Ethereum.
These properties enhance auditability and transparency and improve the overall trustworthiness of the system.Some have conceptualized blockchain as a social trust machine (https://www.economist.com/leaders/2015/10/31/the-trust-machine(accessed on 1 April 2024)).The trust users place in blockchain systems predominantly stems from the cryptographic viewpoint that the majority of miners will act as semi-honest participants.The semi-honest assumption essentially asserts that these miners will respect the predetermined protocols, carrying out actions exactly as specified and programmed in the blockchain software.In particular, the assumption rules out the possibility of collusion among miners to disrupt the regular operations of the blockchain.
There is a vast body of literature on utilizing blockchain to address security issues in FL, more information can be found in recent survey papers, such as [20,21].

EIFFeL Framework and Security Analysis
The EIFFeL framework [16] presents an innovative FL approach that effectively tackles the challenges of data privacy and integrity.Moving beyond the conventional serverclient architecture, as outlined in Section 2.1, EIFFeL incorporates a public bulletin board, which broadcasts intermediary information for secure aggregation and integrity checks.This unique architecture positions all clients as verifiers for each other, with the server playing a collective role.In more detail, secure aggregation enables each client to conceal its parameter update within the aggregated value, by leveraging secret sharing and encryption techniques.The integrity property is ensured through non-interactive proof techniques and the adoption of a public bulletin board.In summary, EIFFeL is designed to tolerate multiple malicious client interventions while ensuring the proper aggregation of model parameters.

Threat Model
In EIFFeL, it assumes that all honest clients correctly follow the protocol and have properly structured inputs.(2) mark the inputs from malicious clients as valid, so as to decide which one will be aggregated.

Scheme Architecture and Workflow
As mentioned previously, EIFFeL consists of a public bulletin board B, a single server S, and n clients, C i (1 ≤ i ≤ n).It is designed to tolerate a limited number of malicious m < n 3 .A high-level sketch of the interactions among a client and other entities is shown in Figure 2 while a detailed description of the EIFFeL scheme is shown in Figures 3 and 4. Note that, for simplicity, we only show one client in the diagram.The numbering of the iteration steps corresponds to the description in Figure 4.In order to ensure (1) privacy for all honest clients, and (2) input integrity, where the server is motivated to verify if all individual updates are well-formed, all clients function as verifiers for each other, while the server also participates collectively in this role.Verification remains achievable even in the presence of m malicious clients.To achieve the designated security guarantees, EIFFeL relies on several cryptographic building blocks (we omit the detailed definition of algorithms; refer to [16] for the full definitions), with each serving a specific purpose in enhancing the scheme's security: -Shamir's t-out-of-n secret sharing scheme [22]: This cryptographic method facilitates the distribution of a secret among n participants, and it requires at least t participants (the threshold) to collectively reconstruct the original secret.Secret-shared non-interactive proofs (SNIPs) [24]: These are cryptographic protocols that allow multiple parties to jointly prove the truth of a statement without revealing their individual inputs.These proofs are constructed in such a way that the validity of the statement can be verified without requiring interaction between the parties.A public validation predicate Valid(•) is defined to conduct the integrity check.
The EIFFeL framework includes a setup phase and a model training phase.In the model training phase, each training epoch consists of four primary steps, as summarized in Figure 1: (1) local model training, (2) model parameter sharing, (3) model parameter aggregation, and (4) model update.We briefly recap the set and training phases in Figure 3 and Figure 4, respectively, which visually encapsulate the essence of its mechanism. -Setup: • All parties are given the security parameter, k, the number of clients, n, the threshold for malicious clients, m, and a field F for secret sharing usage.All clients honestly generate pp ← KA.gen(k), and the server, S, initializes the malicious client list, C * = ∅, and lists Flag[i] = ∅ for clients that have flagged the client, C i , as malicious.

Model parameter sharing:
The following steps are performed: (a).Each client, C i , generates its protected parameters as follows: i.
Establishes n − 1 common secret keys sk i,j ← KA.agree(pk j , sk i ) with each of the other clients. ii.
Generates a proof Creates shares of its update u i for all clients, denoted as {(1, Creates shares of its proof π i for other clients: Publishes the encrypted proof strings, ∀C j ∈ C \i , (j, u j,1 )||(j, π j,1 ) ← AE.enc sk i,j ((j, u i,j )||(j, π i,j )), π i,j = h i,j ||a i,j ||b i,j ||c i,j on the public bulletin board B with its check strings (Ψ u i , Ψ π i ).(b).
The protected parameters are verified as follows: (i) Verifying validity of secret shares: Each client C i : (A).

(B).
Verifies the shares u j,i (π j,i ) using check string If any share fails to be decrypted or verified, flag its creator on B. Server S: (A).
Upon receiving a report (e.g., C i flags C j ), updates Flag Updates C * as follows: For a client C i that has been reported, but with |Flag[i]| ≤ m, server S intervenes for further verification.S requests the shares (in clear) from C i for the clients who flagged it (e.g., ((j, u i,j )||(j, π i,j )) generated by C i for C j ), and verifies the contents with the relevant check string (e.g., (Ψ u i , Ψ π i )).If the verification fails, C i is marked as malicious; otherwise, C j is instructed to use the released share for its computations.

(C).
Announces the updated C * on the public bulletin board B. (ii) Generation of proof summaries by the clients: Server S: -Announces a random number, r ∈ F, on the public bulletin board, B.
Each client C i : -Generates a summary, σ j,i , of the proof string, π j,i , based on r and SNIP, ∀C j / ∈ C * , and publishes it on the public bulletin, B.

(iii)
Verification of proof summaries by the server: Server S: -Collects and verifies all proof summaries from C \ C * with robustRecon(•), and updates C * based on the result.
(iv) Each client C i : -If C i ∈ C * , it can initiate a dispute by transmitting the transcript of the reconstruction of σ i .If any successful dispute occurs, all clients abort the protocol, since the server, S, is considered malicious by withholding the valid updates.-Otherwise, it sends the aggregate U i = ∑ C j / ∈C * u j,i to the server S.

Analysis of the EIFFeL Framework
In [16], the authors performed an analysis of the EIFFeL framework and showed that the framework achieves the pre-defined privacy and integrity properties.In the following, we analyze the assumptions made in their analysis and also demonstrate several observations on the design of the framework.
First of all, the existence of the required public bulletin board and the associated assumptions is very tricky.To guarantee the security of EIFFeL, the public bulletin board should offer the following guarantees: -It should be corrupt-resistant against all entities, including the server, clients, and other attackers.The data should not be changed, deleted, or manipulated by malicious entities in any manner.-It should be able to validate the identities of clients (and the server) according to some public key infrastructure (PKI).By default, some PKI information should be stored on the bulletin board so that it can validate the identity claims of its users.-It should be able to establish a secure channel with its users so that the integrity and authenticity of the users' data can be guaranteed during the transmission.
Public bulletin boards have been used in many scenarios [25], and they can certainly play an important role in FL.However, how to achieve the last two guarantees is not trivial.In particular, the involvement of a PKI means that there will be another fully trusted third party.The trust relationship among all the entities in the solution needs to be clarified in order to guarantee that all threats are properly present.
Secondly, there is some ambiguity in remark 1 of reference [16].It states that EIFFeL prevents the server from "Mark the input of an honest client as invalid and include it in the final aggregate".In fact, the server can manipulate the final aggregate without being noticed, in step 3 of the model parameter aggregation.This is equivalent to invalidating the input of honest clients, as the final aggregate no longer appears to be an aggregation of inputs from all honest clients.This implies that EIFFeL cannot guarantee the integrity of the aggregate in the presence of a malicious server.Note that this observation does not conflict with the desired security properties, i.e., against the server, only privacy is expected.In addition, if the server deviates from the protocol, Lemma 4 from [16] will not hold.This lemma states that "The final aggregate must contain the updates of all honest clients or the protocol is aborted".In fact, the server can maliciously change the aggregate without being detected, as we said before.Although this observation does not show a security flaw in the EIFFeL framework, we regard it as a drawback, and ideally, we should avoid such potential "attacks".This is also motivated by some recent findings (e.g., [26]), which show that failure of integrity (e.g., data poisoning) can lead to privacy breaches.
Thirdly, there are two inefficient designs in the solution.One is about the parameter generation in Step 1 and the key agreement at the beginning of Step 2. During the FL process, all four steps will be iterated hundreds of times in order to reach convergence.According to the description, the key materials need to be repeated in each iteration.This unnecessarily increases the complexity of the solution.These parameters can be set up in the setup phase so that the key materials can be used in all the training iterations (or, epochs).The other involves the usage of authenticated encryption.In the solution, it is assumed that the public bulletin board does not allow any manipulation of the stored data; therefore, the encrypted data in Step 2 will not be manipulated regardless of whether the encryption is authenticated or not.Therefore, standard symmetric encryption will suffice.
Fourthly, the result of the server's operation in Step 2 is confusing.For client C i , which has been reported but |Flag[i]| ≤ m, server S intervenes for further verification.S requests the shares (in clear) from C i for the clients who flagged it (e.g., ((j, u i,j )||(j, π i,j )) generated by C i for C j ), and verifies the contents with the relevant check string (e.g., (Ψ u i , Ψ π i )).If the verification fails, C i is marked as malicious; otherwise, C j is instructed to use the released share for its computations.The server's operation occurs because either C i is maliciously flagged C j or C j is sent the wrong ciphertext to C i , on purpose (note that C j can disclose the right plaintext to make the server's verification pass).However, neither party will be determined as malicious according to the protocol.It is unclear why this is the case.

The Enhanced Scheme
In this section, we describe the enhanced scheme that eliminates the need for any thirdparty server or bulletin board by integrating blockchain technology.It is worth emphasizing that this scheme aims to address the existing privacy and security challenges in the EIFFeL framework, particularly those related to the third-party bulletin board.Certainly, in our design, we attempt to minimize the complexity overhead for all the involved entities.
Our scheme eliminates the need for a central server and any other third party, such as a public bulletin board, thereby mitigating the risks associated with the compromise of such third parties.
Similar to EIFFeL, the enhanced scheme assumes a blockchain platform and multiple (n) clients, and it relies on the same building blocks.In each iteration/epoch, the interactions are shown in Figure 5, where the setup phase is shown in Figure 6 and the numbering of steps corresponds to the description in Figure 7.The overall design of the enhanced scheme is agnostic to any blockchain platform; however, the interaction details between clients and the blockchain (i.e., the smart contracts) may differ in the implementation.To keep the generality and simplicity, we skip the details in our description.The design of smart contracts for the proposed scheme is quite straightforward based on the following fact: when invoking a smart contract, the requester (i.e., a client) can pass the location of necessary input parameters (i.e., where these data are located on the blockchain) so that the smart contract can fetch the data and perform the desired operations in a deterministic manner.To start training, the setup phase is depicted in Figure 6, and the training epoch, shown in Figure 7, will be iterated until a satisfactory global model is achieved.By its nature, information stored on a blockchain is not supposed to be updated.In our description in Figure 7, if a piece of information is stored in the blockchain, then it means that a transaction is made to the blockchain to store the information.When we say that a piece of information is updated (e.g., C * ), we simply mean that a new version of this information is stored on the blockchain, and all follow-up computations should be based on this new version of information. -Setup: • Prepare all the necessary smart contracts for the chosen blockchain.

•
Generate a security parameter, k, the number of clients, n, the threshold for malicious clients, m, and a filed F for secret sharing usage.Initialize the malicious client list, C * = ∅, and lists Flag[i] = ∅ of clients that have flagged the client, C i , as malicious.A validation predicate, Valid(•), is chosen to guarantee integrity, and a collision-resistant hash function H is chosen.All the parameters are stored on the blockchain.

•
A group, G, and its generator, g, are generated for the Burmester-Desmedt group key agreement protocol (Burmester et al., 2005); G and g are stored on the blockchain.

•
Generate pp ← KA.gen(k) for a two-party key agreement scheme, KA.Based on pp, each client, C i , generates its key pair (pk i , sk i ) $ ← KA.gen(pp) and stores the public key on the blockchain.Note that pp is also stored on the blockchain.Choose a standard symmetric key encryption scheme, SE (e.g., AES in the counter mode), and store its information on the blockchain.

•
Each client, C i , establishes n − 1 common secret keys sk i,j ← KA.agree(pk j , sk i ) with every other client.It stores the hash values H(sk i,j ||i||j) (1 ≤ j ̸ = i ≤ n) on the blockchain.

•
All clients check the hash values of secret keys generated by others and resolve any mistakes, if there are any.Note that sk i,j can be computed by both clients, C i , and C j , so that they can check each other's hash values.
should be tamper-resistant and will not collude with any entity in the FL system (i.e., none of the clients should be able to collude with the blockchain).For simplicity, we assume there is a secure channel between the blockchain and any client, in the sense that the client will be authenticated for its access to the blockchain, and the communication is protected regarding its integrity.Similar to EIFFeL, we assume that, at most, m < n 3 malicious users are tolerated, and the threshold of Shamir's t-out-of-n secret sharing scheme is also set as t = m + 1.We omit attacks with the single purpose of denial of service (DoS).In practice, such attacks can be prevented by additional security mechanisms.
In comparison to the EIFFeL scheme, the major philosophy of the enhanced scheme is to use the blockchain platform to replace both the public bulletin board and the server.To this end, all the actions performed by the blockchain platform are deterministic.Apart from the new setup phase, the following major changes have been made: (1) regarding the fourth comment in Section 3.3, a new verification procedure has been proposed for the blockchain platform to trace the malicious client.(2) Leveraging the blockchain platform, the clients execute a Burmester-Desmedt group key agreement protocol [27] to generate a shared group key.The validity of the key is checked as well.(3) Instead of asking a third party to compute the aggregate, the clients store the encrypted shares under the group key on the blockchain platform, so that they can recover the aggregate locally.In contrast to the EIFFeL scheme, the need for a dedicated server has been eliminated.As a result, the risks following the compromise of the server are also eliminated.By asking the clients to share the credentials and intermediary results on the blockchain platform, the clients can check by themselves the validity of the credentials and results from other users.This is facilitated by making the validation operations deterministic.
It is worth stressing that the enhanced scheme preserves all the privacy and integrity properties of the EIFFeL scheme due to the fact that all the security and integrity mechanisms have been kept in the new scheme.

Performance Analysis
We implemented the enhanced scheme in Python with several existing libraries, such as NumPy [28] and Cryptography [29] (the source code of the enhanced scheme instantiation is available at https://github.com/MoienBowen/Blockchain-Federated-Learning(accessed on 1 April 2024)).In the experiment, we used a PC as the simulation platform, with an Intel ® Core™i7-4770 CPU @ 3.4 GHz processor with 16 GB RAM.
Dataset.Given our primary focus on 5G security, we selected a dataset tailored to this domain: the 5G-NIDD dataset [30], a robust compilation of network intrusion data specifically designed for 5G wireless networks, containing 1,215,890 records.This dataset was meticulously generated on an operational 5G testbed, part of the 5G test network (5GTN) at the University of Oulu, Finland.It features an extensive range of simulated attack scenarios, offering a valuable dataset for AI/ML model training.Detailed categories and their respective distributions are presented in Table 1.
Setup.Our parameter selection was guided by our goal of 128-bit security.For example, we used the Secp256r1 curve for the Burmester-Desmedt group key agreement protocol [27], AES-CTR for encryption, SHA-256 for hashing, and a 256-bit prime field as F.Moreover, we considered that there were n = 50, 100, 150, and 200 clients, tolerating up to m = n 10 malicious ones, with d = 1000 as the size of the update gradient vector.Each client owned 24,053, 12,156, 8100, and 6075 records, respectively (each category of the dataset was divided equally and randomly into n copies).We omitted the consensus time of the blockchain but this did not affect the representation of the protocol performance.Regarding the blockchain in Step 2, we calculate the computational time on the PC we use.In practice, the computational time will be based on the mining nodes.Moreover, the complexity will also come from the consensus protocol used in the blockchain platform.Obtaining precise running time statistics on chosen blockchain platforms will be looked at in future work.

Conclusions
This paper culminates with an in-depth discussion of the enhanced scheme for FL, which harnesses the power of blockchain technology.The scheme eliminates the need for a dedicated server and a public bulletin board in the EIFFeL scheme, thereby reducing the risks associated with the compromise of these entities.The clients share credentials and intermediary results on the blockchain platform, allowing them to verify the validity of credentials and results from other users.This paper provides a performance analysis of the enhanced scheme, demonstrating its efficiency and effectiveness in the realm of FL.This paper shows that blockchain technology can enhance the privacy and integrity of FL and open up new possibilities for collaborative ML in various domains.
There are several directions for future work.Here, we only mention two examples.One is to fully implement the proposed scheme, particularly by selecting a blockchain platform, and encoding the desired operations into smart contracts.This will help us understand the overall complexity of the proposed scheme.The other one is to incorporate the concept of differential privacy to further reduce the privacy risks.To this end, it

Figure 2 .
Figure 2. System Architecture of the EIFFeL Scheme.

3 . 4 .
Model parameter aggregation: Server S recovers the final aggregate, U ← robustRecon({i, U i )} C j / ∈C * , and sends it to clients.Model update: Each client, C i , resets its local model with the parameters, U .

Figure 4 .
Figure 4.One iteration/epoch of the EIFFeL scheme.
KA.gen(pp)and announces the public key on the public bulletin board B.
1. Local model training:Each client C i :-Generates its key pair (pk i , sk i ) $ ← Server S: -Publishes the validation predicate Valid(•) on the public bulletin board B.

Table 1 .
The 5G-NIDD dataset attack categories and distribution.Results.In Table 2, we emphasize the runtime of Step 2-model parameter sharing andStep 3-model parameter aggregation, as the performance mainly depends on the number of clients.The values represent the average execution times of single iterations/epochs over 10 runs.

Table 2 .
Runtime of proposed scheme in ms.