Poster: Towards Explainable Artificial Intelligence for Network

Function Virtualization
Sachin Sharma, Avishek Nag, Luis Cordeiro, Omran Ayoub, Massimo Tornatore, Maziar Nekovee
National College of Ireland, UCD, OneSource Portugal, Politecnico di Milano, Italy, University of Sussex
[email protected]
ABSTRACT the network function, and are known as ‘monoliths’. Further, some
Network Function Virtualization (NFV) refers to the process of telecom service providers e.g., BT and Telefonica are now look-
running network functions in virtualized IT infrastructures as soft- ing forward to the 𝜇services-based architecture for their VNFs1 .
warized Virtual Network Functions (VNFs). Several telecom service The 𝜇services-based architecture de-aggregates a large application
providers are currently benefiting from this concept, as it enables into its sub-components (i.e., 𝜇services) and deploys them in a net-
a faster introduction of new network services, thereby meeting work. This facilitates a more flexible, lightweight system, as smaller
changing requirements. Following a trend initially adopted by cloud components are easier to process.
service providers, telecom service providers are also adopting de- The several 𝜇services de-aggregated from a VNF (or a VNF ser-
aggregation of the VNFs into microservices (𝜇services). However, vice chain) have complex mesh-like inter-dependencies among
a 𝜇service-based architecture that can manage a large set of di- them [3, 5] i.e., they may require to call each other several times to
verse and sensitive network functions requires new Artificial Intel- set up a network service. Hence, catering to these inter-dependencies,
ligence (AI)-based methodologies to cope with the complexity of which are typically governed by strict network requirements in
the 𝜇service-based NFV paradigm. This paper focuses on the use of terms of traffic volume, latency, quality of service, etc., makes the
explainable AI (XAI) for gradually migrating towards a 𝜇services- 𝜇services management unprecedentedly complex. Therefore, we
based architecture in NFV. The paper first establishes the need for need an Artificial Intelligence (AI) framework to solve the above
XAI to transform the NFV architecture to a µservice-based architec- problems where such inter-dependencies can be modelled using the
ture and then describes some of our research objectives. Afterwards, data generated from the network. Furthermore, the users of these
our preliminary approach and long-term visions are provided. models (e.g., network operators, cloud providers, etc.) need to know
why a specific AI model arrived at a specific decision in a specific
CCS CONCEPTS scenario, and this requires novel techniques from the explainable
AI (XAI) field. This paper focuses on XAI for 𝜇services in NFV (a
• CCS->Networks->Network architectures->Network design
new field of study for NFV). It presents the next steps in the field
principles ;
of XAI in NFV, describes our envisioned XAI-based 𝜇services de-
aggregation framework (mapped with the NFV-MANO standard)
KEYWORDS
and gives our long-term vision.
Network Function Virtualization, explainable AI, Microservices
ACM Reference Format: 2 XAI-BASED 𝜇SERVICES DE-AGGREGATION
Sachin Sharma, Avishek Nag, Luis Cordeiro, Omran Ayoub, Massimo Tor-
natore, Maziar Nekovee. 2020. Poster: Towards Explainable Artificial In-
telligence for Network Function Virtualization. In The 16th International User Layer μ μ Service Legend:
Requirements VNF1 VNF2 VNF3 μ = Microservice
Conference on emerging Networking EXperiments and Technologies (CoNEXT μ Layer NS = Network Service
XAI μ μ UUID = Universal Unique
’20), December 1–4, 2020, Barcelona, Spain. ACM, New York, NY, USA, 2 pages. Microservice Identification Number
μ VNF/NS
https://doi.org/10.1145/3386367.3431673 Graph
Descriptors
VIM = Virtualized
Infrastructure Manager
(with XAI) XAI Engine
Con = Container or VM
Orchestration Layer
1 INTRODUCTION NFV MANO
Monitoring AI Layer Aggregation/De
Future telecom networks will be software-defined [1], with most Framework AI Prediction Engine Service
-aggregation
Abstraction
network functions virtualized and deployed as software running on Classification Tools
Prediction Resource <VNF+NS UUID>
virtual machines in cloud data centres [4]. These Virtual Network Monitoring
Tools
Abstraction
Failure/Anomaly
Functions (VNFs) in the current mode of operation are deployed Detection
VIM Adaptor Service
as a single piece of software code capturing all functionalities of Mapping
<VM/Con UUID> Database
Infrastructure Layer
Permission to make digital or hard copies of all or part of this work for personal or
Database
classroom use is granted without fee provided that copies are not made or distributed
<VM/Con UUID>
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a Figure 1: Proposed XAI-based 𝜇services De-aggregation
fee. Request permissions from [email protected]. Framework
CoNEXT ’20, December 1–4, 2020, Barcelona, Spain
© 2020 Association for Computing Machinery.
1 https://www.sdxcentral.com/networking/nfv/definitions/microservices-
ACM ISBN 978-1-4503-7948-9/20/12. . . $15.00
https://doi.org/10.1145/3386367.3431673 architecture-telco-cloud
CoNEXT ’20, December 1–4, 2020, Barcelona, Spain S. Sharma, et al.

Fig. 1 shows integration of XAI in NFV MANO architecture results, evaluation); and considering social and ethical aspects too,
(our proposed XAI-based 𝜇services de-aggregation framework). b) Introducing explainability in the AI lifecycle, c) Validating and
In Fig. 1, we have user, service, orchestration and infrastructure evaluating explainability; i.e., the capability to answer questions as:
layers. There is also a monitoring framework (to collect data and are models and outcomes explained effectively to all stakeholders?
monitor the network) and service-mapping framework (to store the Are transparency and trust enhanced?, (3) Measuring effectiveness
mapping of VMs). The AI layer will run in parallel to NFV-MANO of the explanations, e.g., transparency, trust, and usability, Ensur-
and the XAI part will be on top of it. The AI layer will provide the ing interpretability of the models and (4) Getting results in real
decisions of 𝜇services de-aggregation and the XAI part will provide settings using testbed evaluations to mimic the settings of practical
explainability. networks.
First and foremost, to effectively apply AI to any problem, data As the number of state-of-art XAI frameworks is very limited
availability is crucial. The above issue is even more important in ap- (e.g., for measuring transparency and trust), these frameworks need
plying AI techniques to solve telecom-network problems. This is be- to be adapted substantially to meet above objectives.
cause data from real operational networks are difficult to obtain but
may lead to inaccurate results in terms of training the AI algorithms. 4 LONG TERM VISION
A sub-optimal arrangement can be the usage of data generated from AI systems are mainly composed of machine learning, deep learn-
publicly available testbeds like Fed4Fire+https://www.fed4fire.eu/. ing and statistical tools. Although AI is already used in network
Starting with these data, known XAI frameworks such as, e.g., LIME industry, it is not typically used in real time operations. In order to
[2] (Local Interpretable Model-Agnostic Explanations) and SHAP move to real-time operations, network operators need systems that
(SHapley Additive exPlanations) can be applied to develop intelli- are capable of learning automatically. This might be possible with
gent XAI approaches tailored to our problem. The adoption of a AI somehow integrated with the next-generation 𝜇services-based
specific XAI framework depends on the considered surrogate AI NFV. Network automation platforms, such as the Open Networking
model (i.e., the main AI algorithm for the decision-making process). Automation Platform (ONAP), may add AI techniques to support
As the number of XAI frameworks is very limited with their such visions.
own drawbacks, existing XAI tools needs to be adapted to suit the As was also pointed out by a number of executives from the
proposed research problem and address specific requirements not Telecom industry at the 2019 AI Telco Summit3 , there is serious
fully addressed by current models. To highlight the added value concern about AI being a “black box” or the lack of explainability.
that explainability caters, a comparison of the performance of the The development of novel scientific methods to understand how or
proposed XAI algorithms to that of non-explainable AI algorithms why an AI-enabled telecommunication system has provided a spe-
is essential. This comparison will allow us to quantify the benefits cific output can advance and accelerate the vision of a completely
of explainability in our problem and, more importantly, highlight autonomous network. XAI is not only a key requirement for net-
the effectiveness of XAI in communication networks. Moreover, work operators to ensure that their automated systems are working
linear and non-linear optimization techniques will also be employed. as expected, but it might also be necessary to meet requirements, or
Such optimization techniques serve as ground truth to validate it might be important in changing environment where the require-
proposed XAI algorithms, such as in the problem of decomposition ments change very quickly. There can also be challenges involved
and placement of 𝜇services, and can also be used along with the in creating explainable AI in telecommunication system: there may
XAI algorithms. be concerns about user/organization privacy or to put in-place
controls to ensure that the explanations provided are according
3 NEXT STEPS AND CHALLENGES to organization/user privacy. Further, the explanations should be
Starting directly with a 𝜇services architecture is not always prag- correct and reliable. XAI-based 𝜇services-deaggregation framework
matic and at times can be risky too2 . In fact, VNFs should be de- will ensure accountability across the full analytics pipeline – from
ployed as monoliths and only when the system complexity in- data collection to decision and also take into account the needs of
creases, one should start ‘peeling off’ 𝜇services from the monoliths. different people working with the system, considering what differnt
The decision about system complexity and 𝜇services de-aggregation types of explanation might be useful and for what purpose.
can be taken by an AI framework. Furthermore, considering the
fundamental roles that 𝜇services-based applications will play in REFERENCES
our future society (e.g., orchestrating a connection establishment [1] D. Kreutz et. al. 2015. Software-Defined Networking: A Comprehensive Survey.
Proc. IEEE 103, 1 (2015), 14–76.
for a self-driving car or for a remote surgery), the AI models for [2] M.Ribeiro et al. 2016. "Why Should I Trust You?": Explaining the Predictions of
𝜇services management should be explainable, reproducible, and Any Classifier (KDD ’16). ACM-KDD, New York, NY, USA, 1135–1144.
trustworthy. New XAI-based solutions for 𝜇services de-aggregation [3] M. Nekovee et. al. 2020. Towards AI-enabled Microservice Architecture for Net-
work Function Virtualization. IEEE COMNET (2020).
require research breakthroughs that go well beyond the current [4] R. Mijumbi et. al. 2016. Network Function Virtualization: State-of-the-Art and
state-of-the-art in this area. Research Challenges. IEEE Communications Surveys Tutorials 18, 1 (2016), 236–262.
[5] S. Sharma et. al. 2019. On Monolithic and Microservice Deployment of Network
The followings are the next steps of the XAI work: (1) Analyse use Functions. In IEEE Conference on Network Softwarization (NetSoft). 387–395.
cases, requirements and complexity of the 𝜇services de-aggregation
problem, (2) Devise XAI algorithms for 𝜇service decomposition
by: a) identifying and eliminating biases (data collection, training,
2 https://www.martinfowler.com/bliki/MonolithFirst.html 3 https://tmt.knect365.com/software-driven-operations/