This work presents the joint research activities on AI in and for 6G carried out by University of Pisa, Intel Corporation Italia s.p.a. and Telecom Italia s.p.a., within the Hexa-X EU project. Specifically, we focus on Federated Learning of Explainable Artificial Intelligence (Fed-XAI), which has been recently awarded as key innovation by the EU Innovation Radar. We present the main recent achievements, that can be summarised in algorithms for generating federated XAI models in a privacypreserved environment, a communication framework for Federated-Learning-as-a-Service and orchestration algorithms of federated learning participants. Finally, we discuss a proof of concept, that showcases the aforementioned Fed-XAI components.
The fourth (4G) and fith (5G) generations of cellular networks have not only allowed billions of people to communicate with each other, but they have also supported and boosted the digitalization of industries and public nized by CINI, May 29–31, 2023, Pisa, Italy ∗Corresponding author.
CEUR Workshop Proce dings htp:/ceur-ws.org ISN1613-073
CEUR
Workshop Proceedings (CEUR-WS.org) 1Hexa-X website: www.hexa-x.eu tems [2], considering trustworthiness, inclusiveness and sustainability, as its three core values of next generation communication networks [3]. Trustworthiness requirements play a key role in Hexa-X as B5G/6G networks are expected to support a plethora of applications based on artificial intelligence (AI). Indeed, most of these applications will be highly pervasive in the daily processes of individuals, companies and institutions, and also used in critical domains. Therefore, native B5G/6G applications, based on AI models, pose increasing privacy, security and trust issues.
Within the Hexa-X project, the working group composed by University of Pisa, Intel Corporation Italia s.p.a. and Telecom Italia s.p.a. (TIM), collaborates in research activities in the context of federated learning (FL) of exis proposed as a machine learning (ML) paradigm compliant for building trusted and ethical next generation networks. Indeed, on one hand it respects data privacy and on the other hand it ensures the transparency of the AI models [5]. Notably, Fed-XAI has been recently awarded as key innovation by the EU Innovation Radar3.
the development of: i) ad hoc FL strategies for learning FL as a service (FLaaS) on Multi-access Edge Computing 3https://www.innoradar.eu/innovation/45988 gramme2. Hexa-X provides a flagship vision targeted the Hexa-X project1 within the Horizon 2020 pro- plainable AI (XAI) models, denoted as Fed-XAI[4] in the following. Fed-XAI can be regarded as an enabler for to lay the foundations of Beyond 5G (B5G) and 6G sys- several families of use cases envisioned for B5G/6G and © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License XAI models, ii) a communication framework supporting architectures, and iii) orchestration algorithms for FL. the FL literature. However, they are generally considered Moreover, a proof-of-concept will be delivered: a real as opaque, i.e. hard to interpret [6]. Fed-XAI application will be deployed on an edge comput- In certain scenarios other classes of models, such as ing node of an emulated B5G network. Specifically, we Decision Trees (DTs) and Rule-based Systems (RBS), may consider a realistic scenario Vehicle-to-Everything (V2X) be advisable due to their inherent interpretability and [4], in which users collaborates to build an XAI model, in competitive performance, especially when data are repa FL fashion, for performing Quality of Experience (QoE) resented in tabular form. However, they require ad-hoc prediction for video streaming [5]. training strategies for the federated setting since they
The rest of the paper is organized as follows. Section are not typically learned through the optimization of a 2 provides some details on the activities carried out in diferentiable global objective function. Recently, a lot of the Hexa-X project by the research group. In Section research efort in the field of Fed-XAI aims to address this 3, we briefly introduce our proof-of-concept. Section challenge. In the framework of FL of transparent models, 4 argues on the impact that Fed-XAI may have in the Takagi–Sugeno-Kang Fuzzy Rule-Based systems (TSKmarket. Finally, Section 5 draws concluding remarks. FRBSs) have been recently investigated [7]. These regression systems simply consist of collections of rules in the form “IF antecedent THEN consequent”, where the param2. Activities on Fed-XAI in Hexa-X eters (in both the antecedent and consequent parts) are learned from the training data: the antecedent of a rule The enabling of Fed-XAI within the Hexa-X project re- identifies a specific region of the attribute space through quired to carry out the following four diferent and com- highly interpretable linguistic labels, whereas the corplementary activities: responding consequent allows generating the predicted • developing ad-hoc strategies for FL of XAI models, output within such a region as a linear combination of specifically rule-based models; the input variables. In one of our recent works [8] we proposed a novel approach for FL of TSK-FRBS under the • developing a Federated-Learning-as-a-Service orchestration of a central entity. A high level illustration (FLaaS) communication framework; of the approach is depicted in Fig. 1. • developing orchestration strategies for the FLaaS
architecture; • gathering, analyzing and pre-elaborating real
data from live TIM mobile network.
In standard FL a centralised topology is considered: a central entity fulfils the role of orchestrating the process and aggregating local model contributions, coming from the data owners, into a global updated model. In such a setting, FL algorithms mainly aim to optimize a global diferentiable objective function, e.g., through variants of stochastic gradient descent (SGD) such as Federated Averaging (FedAvg) and Federated SGD. Consequently, models like Neural Networks, typically optimized through SGD, are immediately suitable for this type of collaborative optimization and have been widely investigated in
In a nutshell, our approach involves an one-shot communication scheme: each participant learns a TSK-FRBS based on its local and private data, and shares the rule base with the central server. Then, the aggregation procedure consists in the juxtaposition of the rule bases received from the involved participants and in the resolution of rule conflicts. A conflict emerges when two or more rules from diferent participants have the same antecedent and diferent consequents. The aggregation strategy in this case consists in combining the conflicting rules into a single one with the same antecedent: the coeficients of the linear model of the new consequent are evaluated as the average of the coeficients of the original rules, weighted according to the support (how much a rule is activated) and confidence (how good the rule is in modelling data) of each rule as measured on the training data. More details on the proposed FL scheme for TSK-FRBS along with some preliminary results can be found in [8]
Manager (FLGM) per FL process, that is responsible for model aggregation. As already stated, FLLMs may reside directly on the UE or run as MEC Apps. They include the logic to interface with the FLGM, and they may include a learning module, for allowing their owner to participate in the FL process, and an inference module, for making prediction using an available XAI model. The FLGM comprises two modules: the FL process controller (FLPC) and the FL process computation engine (FLPCE). The former manages control-plane interactions with the FLSP (e.g., authorization grants) and the FLLMs, and the latter is the entity that actually builds the global AI model, by exchanging local and global AI model updates with the learning modules of the FLLMs.
We implemented the above FLaaS framework within Simu5G [10, 11], a model library for the OMNeT++ discrete-event simulation framework4, which provides the modules for the simulation of the data plane of 4G/5G mobile networks, as well as for ETSI MEC. In the latter, FL entities are developed as ETSI MEC applications running on a MEC host, with the exception of the FLLM, that can instead be deployed on the UE according to its needs.
allows B5G/6G networks to ofer support for FL applications to their users. The objective is to ofer users (i.e., User Equipments or UEs) the ability to: i) discover FL services and applications that the network may support, ii) join them and therefore use XAI models built in a federated fashion, and iii) contribute to the FL process without sharing their data. The proposed FLaaS is meant to accommodate UEs with high computing power (e.g., cars), that can both acquire data and participate in the FL process, as well as UEs with low computing power (e.g., IoT devices), which can sense but would not be able to participate in the FL process. This last type of devices can leverage Multi-access Edge Computing (MEC) [9], and delegate the computationally onerous task of training and 2.3. Developing Orchestration Strategies exchanging models to MEC Apps. In our FLaaS, network for the FLaaS architecture operators or third parties define onboard FL services.
UEs, on the other hand, have an interface to discover the AI/ML-driven processes are expected to be crucial in manFL services ofered by their network, to join/leave run- aging several aspects of B5G/6G networks, and therefore ning instances of these services, to participate in the FL are being natively included both as elements to be mantraining process, or to just receive model updates. FLaaS aged by the infrastructure and as elements aiding the can run simultaneously several instances of the same FL infrastructure management. One of the Hexa-X activiservice, called FL processes, which are useful to handle ties has focused on this aspect by defining an orchestradiferent sets of UEs in diferent geographical areas. This tion architecture to support the lifecycle of AI/ML-driven also allows the network to orchestrate FL processes ac- processes, enabling their dynamic creation, configuracording to some criteria, for instance the type of UEs, the tion and migration, but also supporting other orchestratype of data that they possess, the geographical location, tion entities to make use of above mentioned processes the computation capabilities, etc. [12]. As anticipated in the previous section, the proposed FLaaS framework support this view thanks to its ability to execute several instances of the same FL processes, simultaneously and independently, thus paving the way for enforcing complex orchestration policies.
Orchestration of AI/ML processes is indeed a critical feature to ensure the feasibility of the proposed Fed-XAI, as an FL process may involve a large number of participants. From a system perspective it is not desirable to include all of them in the same FL instance. Within Figure 2: Overview of the entities involved in the FLaaS this context, several orchestration policies can be conframework sidered [13]. A first type of orchestration strategies may regard the selection of participants to the FL process. Sev
Figure 2 presents a high-level view of the FLaaS frame- eral performance KPIs, such as model accuracy, model work. The FL Service Provider (FLSP) maintains a library explainability and communication-resource consumpof available FL services and the list of running FL pro- tion, may be considered in the selection process. The cesses. An FL Local Manager (FLLM) is associated with each data owner (i.e., UE), and there is one FL Global 4https://omnetpp.org strategies should take into consideration the diferent We consider a ToD scenario, where connected cars trade-ofs between the diferent KPIs, with the aim of send video streams to a remote driver at the edge of the meeting the defined requirements (both functional and network, representing the pilot-seat view. ToD is only non-functional). A second type of orchestration strate- safe if this video stream plays smoothly and with good gies may address the problem of selecting the timing that enough quality. This makes it important to be able to governs FL interactions, such as how often agents should predict loss of quality in advance, in order to allow both report to the federation server or how often the server the network and the remote driver to take the approshould send updates to federated agents. The orches- priate countermeasures. Individual cars participate in tration policies discussed above can only be efectively the collaborative training of an XAI model for predictaddressed if the management of FL tasks is included in ing the quality of the video streaming, under the Feda holistic orchestration system that can i) enforce coor- XAI paradigm and using the FLaaS framework discussed dination among multiple FL tasks to avoid overloading above. The XAI model enables to identify root causes of the underlying communication/computation resources; quality degradation, which is useful for network/system ii) schedule and reserve communication and computation (re)configuration, service level agreement (SLA) verifiresources, possibly working proactively; and iii) moni- cation, and appropriate online reaction on part of the tor the various performance indicators of each FL task, remote driver. For this reason, the Fed-XAI PoC provides checking whether the service requirements are met or a dashboard showing the predictions and root causes in not. real time.
In an initial (ofline) training phase based on the FL 2.4. Data gathering, analysis and learning scheme for TSK-FRBS models discussed in Section 2.1, we exploit Simu5G to generate a training dataset pre-elaboration from TIM live
including a large set of Quality of Service (QoS)-related mobile network data produced by video-streaming sessions. In a subseIn order to test the whole Fed-XAI framework with real- quent (online) prediction phase, Quality of Experience istic B5G/6G data, we gathered live measurements from (QoE) of the video streaming is predicted and the predicTIM’s mobile 5G network, such as base stations positions tions, along with their explanations, are shown on the and user data volume. Moreover, we also acquired geolo- dashboard in real time. The actual video streaming and cated data from live TIM’s Radio Area Network (RAN) by the prediction of its QoE are realized by emulating in real exploiting the Minimization of Drive Test (MDT) func- time a portion of 5G network, which covers roads with tionality. After a preliminary analysis of the extracted semaphore-regulated intersections where connected cars data, they have been pre-elaborated, and aggregated in move. These cars locally run the sender side of the videoorder to use them for feeding Simu5G for generating re- streaming application, whereas the receiver side (that alistic network scenarios for some envisioned B5G/6G represents the remote driver) runs as an application on applications, such as Tele-operated Driving (TOD). The a MEC host. The trafic scenario is set based on data exresulting generated datasets are publicly available5, and tracted from TIM’s live network, extrapolated to model a can be used for building AI models to predict the future future 6G trafic configuration. Moreover, the position of quality of video-streams. The advantage of our data- Base Stations (BSs) in the emulation mirrors the actual driven approach adopted in Fed-XAI is twofold: first, pri- one in the city of Turin. vacy preservation is guaranteed by leveraging FL during Figure 3 shows the scheme of the PoC testbed. It incollaborative training of XAI models, especially suitable cludes the real-time emulated network: the cars as conin heterogeneous B5G/6G scenarios; second, an adequate nected to it as 5G UEs by means of 5G base stations degree of explainability of the models themselves is en- called gNodeBs (gNBs). The network is emulated in real sured, with benefits for industrial customers in terms of time on a desktop machine running Simu5G [14] and high dependability, and for end-users in terms of trust- also the MEC system is included in the emulation. The worthiness. video source and the video player are hosted on two different laptops, connected via Ethernet interfaces to the machine that runs Simu5G. Packets generated by / ad3. Fed-XAI Proof-of-Concept dressed to external interfaces are swapped in real time with Simu5G messages, so that they experience the same Results of the Fed-XAI activities have been also made network treatment that they would in a mobile network. apparent in a Proof of Concept (PoC) within the Hexa- The Fed-XAI application supports the collaborative X project. The PoC shows how XAI models, trained in learning of XAI models and their exploitation for infera federated fashion, are able to predict the quality of a ence purposes and is deployed on a dedicated server. The video stream in a vehicular network context. application includes the following modules: the Fed-XAI 5http://docenti.ing.unipi.it/g.nardini/ai6g_qoe_dataset.html Local Manager can be considered as a UE-related control entity of the FL process; the Learning module performs functionalities (e.g., FL agents enabling QoS predictions) the training stage of local XAI models; the Fed-XAI Com- and expose them to their customers (including car OEMs, putation Engine aggregates the local models and pro- but also application developers and system integrators); duces a federated model; the Inference module allows OEMs can also benefit from more information on netthe end-user to leverage an XAI model for performing work predictions, exploitable to improve the automated inference on local data; the Fed-XAI Dashboard allows driving features ofered to their end-customers (i.e., the users to visualize prediction and to leverage explainabil- actual drivers). More in detail, from a business perspecity to uncover model operation. All Fed-XAI application tive, the goal of OEMs is to provide V2X services and modules have been designed and implemented to be com- advanced features related to automated driving that can pliant with an edge computing environment: we exploit leverage the functionalities of communication networks. a fully virtualized architecture and deploy each module For such advanced functionalities, the prediction of the inside a container of the Docker ecosystem6 so as to al- levels of QoS and QoE are critical to assess the reliability low portability regardless of the underlying hardware of the network and of the ofered functionalities themand software infrastructure and for easier migration in selves. On the other hand, MNOs need to provide the real-world mobility scenarios [15]. The actual Fed-XAI required performance in a trusted environment, so that process exploits the Intel OpenFL7 library, purposely ex- an agreement with the OEMs can be found. The boundary tended to support FL of inherently interpretable models, between MNOs and OEMs worlds is typically governed such as TSK-FRBS. Finally, the Fed-XAI dashboard has by a set of Service Level Requirements (SLRs), defining inbeen implemented as a web application. deed the terms and conditions of the agreement between these two stakeholders (see 5GAA reports for the V2X cases8). These SLRs are service-specific, and may include 4. Impact of the proposed Fed-XAI minimum throughput, maximum delay, but also availapproach ability and reliability of the guaranteed KPIs. With this in mind, in the transition to 6G, accurate, timely, and exA benefit of using the Fed-XAI approach is the increase of plainable predictions are critical to provide advanced and trust in AI for 6G-enabled services. This has an immedi- very challenging use cases with a time horizon ranging ate business impact for 6G stakeholders. In fact, the V2X from extremely short to long periods. Thus, it is evident applications described above are typical cases where a how Fed-XAI is critical to improve mutual understanding collaboration (and related business agreement) is needed and trust among stakeholders in the 6G ecosystem. between mobile network operators (MNOs), possibly in partnership with edge-computing service providers and car original equipment manufacturers (OEMs). Both car 5. Conclusions OEMs and MNOs can benefit from explanations about XAI models predictions and any consequent decision making: MNOs can provide a more explainable set of 6G
can be regarded as a key enabler towards trustworthy AI in several application domains, including next gen
7https://github.com/intel/openfl
mission through the H2020 projects Hexa-X (Grant no. 101015956); by the PNRR - M4C2 - Investimento 1.3, Partenariato Esteso PE00000013 - “FAIR - Future Artificial Intelligence Research” - Spoke 1 “Human-centered AI”; by the PNRR “Tuscany Health Ecosystem” (THE) (Ecosistemi dell’Innovazione) - Spoke 6 - Precision Medicine & Personalized Healthcare (CUP I53C22000780001) under the NextGeneration EU programme; by the Italian Ministry of University and Research (MIUR), in the framework of the FoReLab project (Departments of Excellence); and by the Center for Logistic Systems of Livorno. eration wireless networks. In fact, it ensures privacy preservation during the model training stage and explainability through the adoption of highly interpretable models. The study of Fed-XAI is in its early stages: in this paper we have reported on three dimensions which are being investigated in the context of the Hexa-X EU flagship project: the design of ad-hoc strategies for FL of XAI models; the proposal of a communication framework supporting FLaaS on MEC architectures; the design of orchestration algorithms for FL. A Proof-of-Concept is also described, aimed at showcasing the adoption of FedXAI technologies in a vehicular networking application.
Finally, the potential business impact of the adoption of such paradigm is discussed with reference to the same vehicular scenario. of Explainable AI Models in 6G Systems: Towards Secure and Automated Vehicle Networking, Information 13 (2022) 395. [5] J. L. C. Bárcena, M. Daole, P. Ducange, F. Marcelloni,
A. Renda, F. Rufini, A. Schiavo, Fed-XAI: Federated Learning of Explainable Artificial Intelligence Models, in: 3rd Italian Workshop on Explainable
Artificial Intelligence (XAI.it 2022), 2022. [6] A. B. Arrieta, N. Díaz-Rodríguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. García, S. Gil-López, D. Molina, R. Benjamins, et al., Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible
AI, Information fusion 58 (2020) 82–115. [7] X. Zhu, D. Wang, W. Pedrycz, Z. Li, Horizontal
Federated Learning of Takagi–Sugeno Fuzzy RuleBased Models, IEEE Transactions on Fuzzy Systems 30 (2022) 3537–3547. [8] J. L. C. Bárcena, P. Ducange, A. Ercolani, F. Marcelloni, A. Renda, An Approach to Federated Learning of Explainable Fuzzy Regression Models, in: 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), IEEE, 2022, pp. 1–8. [9] F. Giust, X. Costa-Perez, A. Reznik, Multi-access edge computing: An overview of ETSI MEC ISG,
IEEE 5G Tech Focus 1 (2017) 4. [10] G. Nardini, D. Sabella, G. Stea, P. Thakkar, A. Virdis,
Simu5G–An OMNeT++ Library for End-to-End Performance Evaluation of 5G Networks, IEEE Access 8 (2020) 181176–181191. [11] A. Noferi, G. Nardini, G. Stea, A. Virdis, Rapid prototyping and performance evaluation of etsi mec-based applications, Simulation Modelling Practice and Theory 123 (2023) 102700. doi:h t t p s : [1] M. A. Uusitalo, P. Rugeland, M. R. Boldi, E. C. Stri- / / d o i . o r g / 1 0 . 1 0 1 6 / j . s i m p a t . 2 0 2 2 . 1 0 2 7 0 0 . nati, P. Demestichas, M. Ericson, G. P. Fettweis, [12] J. Péerez-Valero, A. Virdis, A. G. Sánchez, M. C. Filippou, A. Gati, M.-H. Hamon, et al., 6G C. Ntogkas, P. Serrano, G. Landi, S. Kukliński, vision, value, use cases and technologies from eu- C. Morin, I. L. Pavón, B. Sayadi, AI-driven Orropean 6G flagship project Hexa-X, IEEE Access 9 chestration for 6G Networking: the Hexa-X vision, (2021) 160004–160020. in: 2022 IEEE Globecom Workshops (GC Wkshps), [2] M. A. Uusitalo, M. Ericson, B. Richerzhagen, E. U. 2022, pp. 1335–1340.
Soykan, P. Rugeland, G. Fettweis, D. Sabella, G. Wik- [13] Hexa-X Deliverable D6.2 - Design of service ström, M. Boldi, M.-H. Hamon, et al., Hexa-X the management and orchestration functionalities, European 6G flagship project, in: 2021 Joint Euro- 2022. URL: hexa-x.eu/wp-content/uploads/2022/05/ pean Conf. on Networks and Communications & Hexa-X_D6.2_V1.1.pdf. 6G Summit (EuCNC/6G Summit), IEEE, 2021, pp. [14] G. Nardini, G. Stea, A. Virdis, Scalable Real-Time 580–585. Emulation of 5G Networks With Simu5G, IEEE [3] Hexa-X Deliverable D1.2 - Expanded 6G vision, Access 9 (2021) 148504–148520. use cases and societal values – including aspects [15] L. Ma, S. Yi, Q. Li, Eficient Service Handof across of sustainability, security and spectrum, 2021. Edge Servers via Docker Container Migration, in: URL: https://hexa-x.eu/wp-content/uploads/2021/ Proceedings of the Second ACM/IEEE Symposium 05/Hexa-X_D1.2.pdf. on Edge Computing, SEC ’17, Association for Com[4] A. Renda, P. Ducange, F. Marcelloni, D. Sabella, puting Machinery, New York, NY, USA, 2017.
M. C. Filippou, G. Nardini, G. Stea, A. Virdis, D. Micheli, D. Rapone, et al., Federated Learning