Research Challenges in Coupling Artificial Intelligence and Network Management

5.1. Problem type and mapping

An increasing number of different AI techniques have been proposed and applied successfully to a growing variety of different problems in different domains, including network management [Mus18], [Xie18]. Some of the most recently proposed AI approaches are clearly advancements of older approaches, which they supersede. Many other AI approaches are not predecessors or successors but simply complementary because they are useful for different problems or optimize different metrics. In fact, different AI approaches are useful for different kinds of problem inputs (e.g., tabular data vs. text vs. images vs. time series) and also for different kinds of desired outputs (e.g., a predicted value, a classification, or an action). Similarly, there may be trade-offs between multiple approaches that take the same kind of inputs and desired outputs (e.g., in terms of desired objective, computation complexity, constraints).¶

Overall, it is a key challenge of using AI techniques for network management to properly understand and map which kind of problems with which inputs, outputs, and objectives are best solved with which kind of AI (or non-AI) approaches. Given the wealth of existing and newly released AI approaches, this is far from a trivial task.¶

5.2. Performance of produced models

From a general point of view, any AI technique will produce results with a certain level of quality. This leads to two inherent questions: (1) what is the definition of the performance in a context of a NM application? (2) How to measure it? (3) How to ensure the quality of produced results by AI is aligned with NM objectives? (4) How to maintain or improve the quality of produced results?¶

Many metrics have been already defined to evaluate the performance of an AI-based technique according to its NM-level objectives. For example, QoS metrics (throughput, latency) can serve to measure the performance of a routing algorithm along with the computational complexity (memory consumption, size of routing tables). The question is to model and measure these two antagonist types of metrics. Number of true/false positives/negatives are the most basic metrics for network attack detection functions. Although the first two questions are thus already answered even if improvement can be done, question (3) refers to the integration of metrics into AI algorithms. Its objective is to obtain the best results which need to be quantified with these metrics. Depending on the type of algorithm, these metrics are either evaluated in an online manner with a feedback loop (for example with reinforcement learning) or in batch to optimize a model based on a particular context (for example described by a dataset for machine learning).¶

The problem is twofold. First, the performance can be measured through multiple metrics of different types (numerical or ordinal, for example), and some can be constrained by fixed boundaries (such as maximum latency), making their joint use challenging when creating an AI model to resolve an NM problem. Second, the scale of the metrics differs in importance or impact and can eventually vary across domains. It can be hard to precisely assess what a good or bad value is (as it might depend on multiple other ones) and it is even more difficult to integrate this in an AI technique, especially for learning algorithms to adjust their models based on performance. Indeed, many learning algorithms run through multiple iterations and rely on internal metrics, Mean Absolute Error (MAE) or Mean Squared Error (MSE) for neural network, Gini index or entropy for decision trees, distance to a hyperplane for SVMs, etc) which are not strongly correlated to the final operational metrics of the NM application. AI-internal metrics such as the loss do not match well the metrics related to the final NM objectives, thus the significance and impact of the AI errors cannot be easily translated into the NM domain.¶

For instance, a decision tree algorithm for classification purposes aims to create branches with a maximum of data from the same classes and to avoid mixing classes. It is done thanks to a criterion like the entropy index, but this kind of index does not assume any difference between mixing classes A and B or A and C. Assuming an operational point of view in which A and B mixing in the predictions is not critical, the algorithm should have preferred to mix A and B rather than A and C, even if in the first case it will produce more errors. Therefore, the internal functioning of the AI algorithms should be refined, specifically by defining a particular criterion to replace entropy as a quality measure when separating two branches. It assumes that the final NM objectives are integrated at this stage.¶

Another concrete example is traffic predictors, which aim to forecast traffic demands. They produce an output that is not necessarily simple to interpret and use by, for example, capacity allocation strategies/policies. A traditional traffic prediction that tries to minimize (perfectly symmetric) MAE/MSE treats positive and negative errors identically, hence is agnostic to the diverse meaning (and costs) of under- and over-provisioning. And, such a prediction does not provide any information on, for example, how to dimension resources/capacity to accommodate future demand while avoiding all under-provisioning (which causes service disruption) while minimizing over-provisioning (i.e., wasting resources). In other words, it forces the operator to guess the over-provisioning by taking uninformed safety margins. A more sensible approach is to forecast directly the needed capacity, rather than the traffic [Beg19].¶

While this is just an example, the high-level challenge is devising forecasting models that minimize the correct objective/loss function for the specific NM task at hand (instead of generic MAE/MSE). In this way, the prediction phase becomes an integral part of NM, and not just a limited and difficult-to-use input to it. In ML terms, this maps to solving the loss-metric mismatch in the context of anticipatory NM [Hua19].¶

Another issue for statistical learning (from examples/observations) is mainly about extracting an estimator from a finite set of input-output samples drawn from an unknown probability distribution that should be descriptive enough for unseen/new input data. In this context, online monitoring and error control of the quality/properties of these point estimators (bias, variance, mean squared error, etc.) is critical for dynamic/uncertain network environments. Similar reasoning/challenge applies to interval estimates, i.e., confidence intervals (frequentist) and credible intervals (Bayesian).¶

Finally, question (4) refers to the ability of an AI solution to remain efficient and eventually improve over time [Ans26]. This requires dynamic methods capable to adapt to a changing environment. As already highlighted, the models can be dynamically adjusted based on the errors they produced. In the context of ML, the models can be also updated based on new data, either through a complete re-learning phase, fine-tuning or transfer-learning. This assumes to collect and ingest continuously new data. However, as highlighted in [Sha21b], this type of ML, qualified as online or incremental, raises several challenges when applied to traffic analysis. For example, there is a set of related challenges related to select or discard some data over a time horizon and to label data in real-time. Other challenges are more generic to this ML research area such as class imbalance or concept drift.¶

5.3. Lightweight AI

Network management and operations often need to be performed under strict time constraints, i.e., at line rate, particularly in autonomic or self-driven networks. Locating NM functions as close as possible to where forwarding is achieved is an interesting option to avoid additional delays when these operations are performed remotely, for example in a centralized controller. Besides, forwarding devices may offer available resources to supplement or replace edge resources. For AI coupled with network management, AI tasks can be offloaded in network devices or more generally embedded within the network. Obviously, time-critical tasks are the best candidates to be offloaded within the network. Costly learning tasks should be processed in high-end servers but created models can be deployed, configured, modified and tuned in switches.¶

Recent advances in network programmability facilitate the programming of specific tasks at the data-plane level. P4 [Bos14] is widely used today for many tasks including firewalling [Dat18] or bandwidth management [Che19]. P4 is generally agnostic to specific hardware. It is based on the RMT (Reconfigurable Match Table) architectural model [Bos13] that is generally accepted to be generic enough to represent limited but essential switch architecture components and functionalities. The RMT model allows reconfiguring match-action tables where actions can be usual ones (rewrite some headers, forward, drop...). Actions are thus applied on the packets when they are forwarded. Actions can also be more complex programs with some safeguards: no loops, resistivity, etc. The impact on the program development is huge. For example, real number operations are not available by default while they are widely used in many AI algorithms.¶

In a nutshell, the first challenge to overcome in embedding AI in a network is the hardware's capacity to support AI operations (architectural limitation). Considering software equipment such as a virtual switch simplifies the problem but does not entirely resolve it, as even in that case, strict line-rate requirements limit the type of programs that can be executed. For example, BPF (Berkeley Packet Filter) [Mcc93] programs provide greater control over packet processing in OVS [Cha18] but still have some limitations, as the execution time of these programs is bounded by nature to ensure their termination, an essential requirement assuming the run-to-completion model which permits high throughput.¶

The second challenge of network-embedded AI in the network is allocating enough resources for AI tasks while minimizing impact on other network device tasks such as forwarding, monitoring, and filtering. Approximation and/or optimization of AI tasks are potential directions to help in this area. For instance, many network monitoring proposals rely on sketches with well-tuned implementations for the data-plane [Liu16][Yan18]. However, no general optimized AI-programmable abstraction exists to fit all cases, and proposals are mostly use-case centric. There have been many proposed approaches to develop specific P4 programs for many NM tasks, including involving ML. For each, this requires a specific adaptation [Hau23] with a few attempts to propose generic programs to be reusable or composed as a kind of libraries such as [Zha24] leveraging quantization, [Jos21][Jos22] relying on pre-computed lookup tables of real-value functions or [Swa23] proposing function templates for common operations. Besides, distributed processing is a common technique to distribute the load of a single task between multiple entities. AI task decomposition between network elements, edge servers, or controllers has also been proposed [Gup18].¶

5.4. Distributed AI

Distributed AI assumes different related tasks and components to be distributed across computational and possibly heterogeneous resources. For example, with advances in transfer and Federated Learning (FL), models can be learned, partially shared and combined or data can be also shared to either improve a local or global model. By nature, a network and a networked system is distributed and is thus well adapted to any distributed application. This is exacerbated with the deployment of fog infrastructure mixing network and computational resources. Hence, network management can directly benefit to the distributed network structure to solve its own particular problems, but any other type of AI-based distributed applications also assumes communication technologies to enable interactions between the different entities. This leads to the two sub-challenges described hereafter.¶

5.5. AI for planning of actions

Many tasks in network management revolve around the planning of actions with the purpose of optimizing a network and facilitating the delivery of communication services. For example, communication paths need to be planned and set up in ways that minimize wasted network resources (to optimize cost) while facilitating high network utilization (avoiding bottlenecks and the formation of congestion hotspots) and ensuring resiliency (by making sure that backup paths are not congruent with primary paths). Other examples were mentioned in Section 3.¶

The promise of central control is that decisions can be optimized when made with complete knowledge of relevant context, as opposed to distributed control, which needs to rely on local decisions made with incomplete knowledge while incurring higher overhead to replicate relevant state across multiple systems. However, as the scale of networks and interconnected systems continues to grow, so does the size of the planning task. Many problems are NP-hard. As a result, solutions typically need to rely on heuristics and algorithms that often result in suboptimal outcomes and that are challenging to deploy in a scalable manner [Ahm21].¶

The emergence of Intent-Based Networking (IBN) emphasizes the need for automated planning even further. IBN should allow users (network operators, not end users of communication services) to articulate desired outcomes without the need to specify how to achieve those outcomes. An Intent-Based System is responsible for translating the intent into courses of action that achieve the desired outcomes and that continue to maintain the outcomes over time [RFC9315]. How the necessary courses of action are derived and what planning needs to take place is left open, but that is where the real challenge lies. Solutions that rely on clever algorithms devised by human developers face the same challenges as any other network management tasks.¶

These properties (problems with a clearly defined need, whose solution is faced with exploding search spaces and that today rely on algorithms and heuristics that in many cases result only suboptimal outcomes and significant limitations in scale) make automated planning of actions an ideal candidate for the application of AI-based solutions [Abd24].¶

A much-publicized leap in AI has been the development of AlphaGo [Sil16]. Instead of using AI to merely solve classification problems, AlphaGo has been successful in automatically deriving winning strategies for board games, specifically the game of Go which features a prohibitively large search space that was long thought to put the ability to play Go at a world class level beyond the reach of problems that AI could solve. Among the remarkable aspects of Alpha Go is that it is able to identify winning strategies completely on its own, without needing those strategies to be taught or learned through observations assuming the system is aware of rules.¶

The challenge for AI in network management is hence, where is the equivalent of an Alpha Go that can be applied to network management (and networking) problems? Specifically, better solutions are needed for solutions that automatically derive plans and courses of actions for network optimization and similar NP-hard problems, such as provided today with only limited effectiveness by controllers and management applications.¶

Although AI-based solutions for the automated planning of actions, including the automated identification of courses of action, have not been explored as much as classification [Sal20] problems, the quest for autonomous networks in the last decade and the advent of 5G and B5G have led to a rapid increase in proposed solutions, such as within the context of Zero Touch Management [Cor22].¶

Also, the evaluation of AI algorithms to derive courses of actions is complex. Contrary to game playing, solutions need to be applied in the real world, where actions have real effects and consequences. Different orientations can be envisioned. First, incremental application of AI decisions with small steps can allow careful observation and detection of unexpected effects. This can be complemented with rollback techniques. Second, verification techniques can be leveraged to verify that decisions made by AI are maintained within safety bounds [Xin24]. Third, sandbox environments can be used but they should be representative of the real world. After progress in simulation and emulation, recent research advances lead to the definition of digital twins, which implies tight coupling between a real system and its digital twin to ensure parallel but synchronized execution [Wu21]. Alternatively, transfer learning techniques are another promising area to capitalize on ML models applicable to real world systems, for instance, to learn an intrusion detector that can be instantiated in multiple environments [Sha24][Ans22]. Generally, it is also an open problem to make the use of AI more acceptable, as highlighted in the dedicated section.¶

6.1. Data for AI-based NM solutions

Assuming a network management application, the first problem to address is to define the data to be collected, which will be appropriate to obtain accurate results. This data selection can require defining problem-specific data or features (feature engineering). Furthermore, machine learning algorithms only work as desired when data to be analyzed respects given properties. Many methods rely on vector-based distances, which assume that the data encoded into the vector respects the underlying distance semantic. Taking the first n bytes of a packet as vectors and computing distances accordingly is possible, but does not embed the semantic of the information carried in the headers. A solution to circumvent the problem of network traffic representation is to transform traffic data into a specific format that can be easily handled by NN architectures, for instance by creating images analyzed by convolutional neural networks [Sha21]. Data can also be easily represented as graphs, because topologies or communication graphs would require less adaptation to be given as inputs to GNNs [Jie22]. As an example, graph-based representations are considered practically efficient due to their intrinsic structural similarities with network data, as illustrated in [Bar23] for attack detection.¶

Since data to handle can also be in a schema-free or eventually text-based format, one example could be the automated annotation of management intents provided in unstructured textual format (policies descriptions, specifications) to extract management entities and operations. For that purpose, suitable annotation models need to be built using existing NER (Named Entity Recognition) techniques usually applied for NLP. However, this shall be carefully crafted or specialized for network management (intent) language which indirectly relates back to the challenges of AI techniques for NM specified earlier. Today, with the progress on LLMs, different proposed approaches have emerged as for example in [Mek24][Fua24][Dze23].
Secondly, similar to the problem of mapping AI algorithms with NM problems in section Section 5.1, data to be collected also depends on the NM problem to be solved. The mapping between the data sources and the problem is not straightforward as all dependencies or correlations are not known in advance and some might be be discovered by the AI algorithms themselves [For23]. In addition, the types of data to collect can vary over time to maintain the performance of an AI-based application or to adapt it to a new context when learned models are updated dynamically.¶

Thirdly, the behaviour of any network is not just derived from the events that can be directly observed, such as network traffic overload, but also from events occurring outside the environment of the network. The information provided by the detectors of such kinds of events, e.g. a natural incident (earthquake, storm), can be used to determine the adaptation of the network to avoid potential problems derived from such events. Those can be provided by big data sources as well as sensors of many kinds. The challenge related to this task is to process large amounts of data and associate it with the effects that those events have on the network. It is hard to determine the static and dynamic relation between the data provided by external sources and the specific implications it may have in networks. For instance, the effect of a “flash crowd” detected in an external source may require adapting the network service configuration to support is related workload. The objective is to complement a control-loop, as shown in [Mar18], by including the specific AI engines into the decision components as well as the processes that close the loop, so the AI engine can receive feedback from the network in order to improve its own behaviour.¶

6.2. Data collection

Once defined, the second problem to address is the collection of data. Monitoring frameworks have been developed for many years such as IPFIX [RFC7011] and more recently with SDN-based monitoring solutions [Yu14][Ngu20]. However, going towards more AI for actions in network management supposes also to retrieve more than traffic related information. Actually, configuration information such as topologies, routing tables or security policies have been proven to be relevant in specific scenarios. As a result, many different technologies can be used to retrieve meaningful data. To support improved QoE, monitoring of the application layer is helpful but far from being easy with the heterogeneity of end-user applications and the wide use of encrypted channels. Monitoring techniques need to be reinvented through the definition of new techniques to extract knowledge from raw measurement [Bri19] or by involving end-users with crowd-sourcing [Hir15] and distributed monitoring. Also, the data-mesh concept [Mac22] proposes to classify data into three categories: source-aligned, aggregate and consumer-aligned. Source-aligned data are those related to the same operational domain, and it is important to correlate or aggregate them with higher planes: management-, control- and forwarding plane. An issue is the difference, not only in the nature of data, but in their volumes and their variety. Some may change rapidly over time (for example network traffic) while other may be quite stable (device state).¶

The collecting process requirements depend on the kind of processing. We can distinguish two major classes: batch/offline vs real-time/online processing. In particular, real-time monitoring tools are key in enabling dynamic resource management functions to operate on short reconfiguration cycles. However, maintaining an accurate view of the network state requires a vast amount of information to be collected and processed. While efficient mechanisms that extract raw measurement data at line rate have been recently developed, the processing of collected data is still a costly operation. This involves potentially sampling, evaluating and aggregating a vast amount of state information as a response to a diverse set of monitoring queries, before generating accurate reports. One difficult problem resides also in the availability of data as real-time data from different sources to be aggregated may not arrive at the same time requiring so some buffering techniques. Machine learning methods, e.g. based on regression, can be used to intelligently filter the raw measurements and thus reduce the volume of data to process. For example, in [Tan20] the authors proposed an approach in which the classifiers derived for this purpose (according to measurements on traffic properties) can achieve a threefold improvement in the query processing capability. A residual question is the storage of raw measurements. In fact, predicting the lifetime of data is challenging because their analysis may not be planned and triggered by a particular event (for example, an anomaly or attack). As a result, the provisioning of storage capacity can be hard.¶

In parallel to the continuously increasing dynamic behavior of networks and complexity of traffic, there is a trend towards more user traffic processing customization [RFC8986][Li19]. As a result, fine grained information about network element states is expected and new proposed approaches have emerged to collect on-path data or in-band network telemetry information [Tan20b]. These new approaches have been designed by introducing much flexibility and customization and could be helpful to be used in conjunction with AI applications. However, the seamless coupling of telemetry processes with packet forwarding requires careful definition of solutions to limit the overhead and the impact of the throughput while providing the necessary level of details. This shares commonalities with the lightweight AI challenge.¶

6.3. Usable data

Although all agree on the necessity to have more shared datasets, it is quite uncommon in practice. Data contains private or sensitive information and may not be shared because of the criticality of data (which can be used by ill-intentioned adversaries) or due to laws or regulations, even within the same company. To solve this issue, anonymization techniques [Dij19] can be enhanced to optimize the trade-off between valuable data and sensitive information (potential) leakage or reconstruction. Whatever the final user of data, regulations and laws impose rules on data management with potentially costly impact if they are not respected voluntarily or not. Defining a new monitoring framework should always consider security and privacy aspects, for example to let any user/customer or access/remove its own data with General Data Protection Regulation (GDPR) in EU. The challenge resides here in the capacity of qualifying what is critical or private information and the capacity for an adversary to reconstruct it from other sources of data. Hence AI/ML based solutions will require more data but also more administrative, legal and ethical procedures. Those can last long and so slow down the deployment of a new solution. In addition, this requires interaction with experts from different domains (e.g. AI engineer and a lawyer) for example to ensure by-design privacy in traffic analysis techniques [Wan22]. The integration of these non-technical constraints should be considered when defining new data to be collected or a new technique to collect data. However, knowing the final use of data is most of the time necessary for ethical and legal assessment which assumes that those considerations should be integrated from the early design of new AI-based solutions.¶

For supervised or semi-supervised training, having a labelled dataset is a prerequisite. It constitutes a major challenge as well. While probes exist to collect real network data, those data are typically unlabelled. This limits application of ML to unsupervised learning tasks (learning from data). Because manual labelling is a tedious task, one option is to leverage AI to guide humans. This may also support a better generalization of a learned model. Indeed, an underlying challenge is the genericity or coverage of the datasets. Labels encode values of an objective function, the challenge posed by the design of such tools is tremendous since for involving a M:N relationship: 1 data type may be associated to M objective function values and N data types may be associated to 1 objective function. As a result, most datasets used for research encode a single label for a particular application like attack label for datasets to be used in the context of intrusion detection or application type for network traffic used for classification where the value of a single dataset could be capitalized in several applications. More generally, in the context of intrusion detection, using raw data rather than pre-processed data as it is common in open dataset has been demonstrated to be inefficient [Dub24].¶

Again, researchers need empirical (or at least realistic) datasets to validate their solutions. Unfortunately, as highlighted above, having such data from real deployments for various reasons (business secrets, privacy concerns, concerns that vulnerabilities are revealed by accident, raw unlabelled data, etc.) is tough. Even if such a dataset is available it might not be enough to convincingly validate a new algorithm. Instead of falling back to artificial testbed experiments or simulation, it would be useful to have the capability to generate datasets with characteristics that are not 100% identical but like the characteristics of one or more real datasets. Such synthetic networks can be used to validate new management algorithms, intrusion detection systems, etc. The usage of AI, Generative Adversarial Networks (GANs), in this area [Hui22] is not yet widespread and there are still many concerns that deter researchers, e.g. the fear of leaking sensitive information from the original dataset into the synthetic dataset. Furthermore, a major underlying challenge is to generate data realistic enough. This requires formalizing and integrating network-specific constraints, such as protocol specification, when generating data [Jia24] or using a real testbed guided by AI to create challenging scenarios for validation purposes [Ans25].¶

7.1. Explainability of Network-AI products

A common issue across many ML applications is their lack to provide human-understandable reasoning processes. This means that, after training, the knowledge acquired by ML models is unintelligible to humans. As a result, offering hard guarantees on performance is a very challenging issue. In addition, complex ML models like neural networks -that often have more than hundreds of thousands of parameters- are very hard to debug or troubleshoot in case of failure.¶

While this is a common issue for all applications of AI, many areas work well with uncertainty and the black-box behaviour of AI-based solutions. For instance, users accept an inherent error in recommender systems or computer vision solutions.¶

The networking field has already produced a set of well-established network management algorithms and methods, with clear performance guarantees and troubleshooting mechanisms [Rex06][Kr14]. As such, improving debugging, troubleshooting and guarantees on AI-based solutions for networking is a must. AI researchers and practitioners are devoting large research efforts to improve this aspect of ML models, which is commonly known as explainability [XAI].¶

This set of techniques provides insights and, in some cases, guarantees on the performance and behaviour of ML-based solutions. Understanding such techniques, researching and applying them to network AI is critical for the success of the field.¶

There exist several ML-based methods that are human-understandable, although not widely used today. For instance, [Mar20] shows a method for building anticipation models (prediction) that provide explanations while determining some actions for tuning some parameters of the network. There are other challenges that should be addressed, such as providing explanations for other ML methods that are quite extended. For instance, xNN/SVM models can be accompanied by Digital Twins of the network that are reversely explored to explain some output from the ML model (e.g., xNN/SVM). In this context, there already exist several methods [Zil20][Puj21] that produce human-readable interpretations of trained NN models, by analysing their neural activations on different inputs. It is worth noting that Digital Twins are not considered per se an AI approach; they merely serve to provide a digital representation of a network that can serve as its proxy and offer a layer of indirection between management applications and actual network resources. However, it is conceivable that AI-based management applications can be combined and operate in conjunction with Digital Twin technology, for example to use a Digital Twins as an experimentation sandbox or staging ground for AI-driven applications.)¶

7.2. AI-based products and algorithms in production systems

AI-based network management and optimization algorithms are first trained, then the resulting model is used to produce relevant inferences in operation, either in management or optimization scenarios. A relevant question for the success of AI-based solutions is: where does this training occur?¶

Traditionally, AI-based models have been trained in the same scenario where they operate[Val17][Xu18], this is the customer network. However, this presents critical drawbacks. First, training an AI model for management and operation typically requires generating network configurations and scenarios that can break the network. This is because training requires seeing a broad spectrum of scenarios. Thus, training in production networks is very challenging. Second, customer networks may not be equipped with the monitoring infrastructure required to collect the data used in the training process (e.g., performance metrics). Performing learning directly into a production network is possible assuming imperfect models and the need of several step of refinement before it gets stable. For non-critical management task, such assumption can hold, and additional safe-guarding mechanisms should be considered in order to keep outputs of ML algorithms (such as decisions) within acceptable boundaries.¶

A more sensible approach is to train the AI-based product in a lab, for instance in the vendor’s premises. In the lab, AI models can be trained in a controlled testbed, with any configuration, even ones that break the network. However, the main challenge here arises from the fundamental differences between the lab’s network and the customer networks. For instance, the topology of the lab’s network might be smaller, etc. As a result, there is a need for the learned models to generalize. In this context, generalization means that models should be able to operate in other scenarios not seen during training, with different topologies, routing configurations, scheduling policies, etc. This well-known problem is due to the dependency between training and testing data when benchmarking models.¶

In order to address this generalization problem, multiple complementary approaches are possible:
One approach is training on diverse data that represents large parts of the expected problem space. For example, training with various traffic patterns will help improving the generalization of an AI-based traffic analyser. Another approach is to leverage AI designs or architectures that facilitate generalization. One example is GNNs [gnn1][gnn2]. They are a type of neural networks able to operate and generalize over graphs. Indeed, networks are fundamentally represented as graphs: topology, routing, etc. With GNNs, vendors can train the AI model in a lab with a certain topology and then directly use the resulting model in different customer networks, even with different network topologies. Finally, another approach is Transfer Learning [tl1]. With this technique, the knowledge gained in the lab’s training is used to operate in the customer network. Transfer Learning still requires that some data from the customer is used to re-train and fine-tune the model (e.g., accurate performance measurements). This means that, for each customer network, re-training is required. This may be problematic since it requires added cost and access to customer data.¶

In addition to the challenge of generalizing from training to production environment, there are also challenges in terms of interoperability between different AI approaches and different deployment environments. As mentioned above, AI approaches may be deployed in diverse environments, e.g., for training and production, but also for local development, for testing, and for validation or in different part of the production systems. These environments may differ in available compute resources, network topology, operating systems, cloud providers, etc. (single node machine, single cluster, many distributed clusters, etc.). Deploying the same AI solutions in these different environments can lead to various challenges in terms of interoperability. Common AI frameworks support scaling across networks of different size. Yet, many frameworks are often combined, e.g., for data collection, processing, predictions, validation, etc. Again, ensuring interoperability between these frameworks can be tedious.¶

This shares similarities with problems described in Section 5.4 and particularly emphasizes the need for network environments to provide interfaces and descriptions suitable for AI solutions to be properly instantiated and configured.¶

One approach to address these interoperability challenges is through meta-frameworks that interface with most available AI frameworks. These meta-frameworks provide a higher level of abstraction and often allow seamless deployment across different environments (e.g., on-premises or at different cloud providers) [Mor18].¶

7.3. AI with humans in the loop

Depending on the network management task, AI can automate and replace manual human control, or it can complement human experts and keep them in the loop. Keeping humans in the loop will be an important step of building trust in AI approaches and help ensure the desired outcomes. There are various ways of keeping humans in the loop in the different fields of AI [Wu22], which could be useful for different aspects of network management.¶

In classification tasks (e.g., detecting security breaches, malware or detecting anomalies), trained AI models provide a confidence score in addition to the predicted class. If the confidence is high, the prediction is used directly. If the confidence is too low, a human expert may jump in and make the decision - thereby also providing valuable training data to improve the AI model. Such approaches are already being used in industry, e.g., to automatically label datasets (AWS SageMake). Similar approaches could also be used for other supervised learning tasks, e.g., regression. Still, it is an open challenge to keep humans in the loop in all phases of the learning process.¶

When using RL, e.g. to control service scaling and placement or route traffic flows, the agents typically interact with the environment (i.e., the simulated or real network) completely autonomously without human feedback. However, there is a growing number of approaches to put human experts back into the loop. One approach is offline reinforcement learning, where the training data does not come from the reinforcement learning agent’s own exploration but from pre-recorded traces of human experts (e.g., placement decisions that were made by humans before). Another approach is to reward the RL agent based on human feedback rather than a pre-defined reward function [Lee21]. Again, while there are first promising approaches, more work is required in this area. Overall, it is an open challenge to both leverage the benefits of AI but keep human experts in the loop where it is useful.¶

9.1. AI-based security solutions

The first security consideration refers to the use of AI for NM problem related to security of the managed networked systems. There are multiple scenarios where AI can be leveraged: to perform traffic filtering, to detect anomalies or to decide on target moving defence strategies. In these cases, the performance of the AI algorithms impacts on the security performance (e.g. detection or mitigation effectiveness) like any other non-AI system. However, AI methods generally tend to obfuscate how predictions are made and decisions taken. Explainability of AI is thus highly important and is addressed globally in section Section 7.1.¶

Assuming a ML trained model, there is always an uncertainty regarding reachable performance on the wild once the solution is deployed as it can suffer from a poor generalization due to different reasons. There are two major problems which are well known in the ML field: overfitting or under-fitting. In the first case, the learned model is too specific to the training data while in the second case the model does not infer any valuable knowledge from data. To avoid these issues, hyper-parameter fine-tuning is necessary. For example, the number of iterations is an essential hyper-parameter to be adjusted to learn a neural network model. If it is too low, the learning does not converge to a representative model of the training data (underfitting). With a high value, there is a risk that the model is too close to the training data (overfitting). In general, finding the right hyper-parameters is helpful to find a good ML algorithm configuration. There are different techniques ranging from grid-search to Bayesian optimization falling into the AI area of Hyper-Parameter Optimization (HPO) [Bis23]. This consideration goes beyond the sole problem of hyper-parameters settings but as a full analysis pipeline also assumes the ML algorithm to be selected or the data pre-processing to be configured. This reflects challenges from the AI research area covered by AutoML technique. In this document, this is also referred in section Section 5.1.3 when considered in the context of NM. As highlighted in the aforementioned section, some expertise or area-specific knowledge can help guiding automated configuration processes.¶

Besides, machine learning assumes to have representative training data. The quality of dataset for learning is a vast problem. Additionally, the representation of data needs to be addressed carefully to be properly analysed by AI models, for example with pre-processing techniques to normalize data, balance classes or encode categorical features depending on the type of algorithms. Actually, section Section 6 of this document fully addresses the concerns related to data in regard to NM problems.¶

9.2. Security of AI

Although ensuring a good performance of AI algorithms is already challenging, assuming an attacker aiming at compromising it emphasizes the problem. Adversarial AI and notably adversarial ML have attracted a lot of attention over the last year. Adversarial AI and ML relates to both attack and defences. While this is out of scope of the document, evaluating threats against an ML system before deploying it is an important aspect. This supposes to assess what types of information the attacker can access (training data, trained model, algorithm configuration...) and the performable malicious actions (inject false training data, test the system, poison a model, etc.) to evaluate the magnitude of the impact of possible attacks.¶

For illustration purposes, we refer hereafter to some examples. In the case of an intrusion detector, an attacker may try to poison training data by providing adversarial samples to ensure that the detector will miss-classify the future attacks [Jmi22]. In a white-box approach where the model is known from the attacker, the attack can be carefully crafted to avoid being properly labelled. For instance, packet sizes and timings can be easily modified to bypass ML-based traffic classification system [Nas21]. In a black-box model, the attacker ignores the functioning or training data of the ML systems but can try to infer some information. For example, the attacker can try to reconstruct sensitive information which have been used for training. This type of attack qualified as model inversion [Fre15] raises concern regarding privacy.¶

All these threats are exacerbated in the context of solutions relying on distributed AI involving multiple entities that are not necessarily controlled by the same authority. Once the threats are assessed, solutions need to be developed and deployed which can be either proactive by providing some guarantees regarding the involved entities using authentication or trust mechanisms but also reactive by validating data processing through voting mechanisms or knowledge proofs. Other solutions include defensive techniques to rate limit or filter queries to a particular deployed model. All these examples are for illustration purposes and are not exhaustive.¶

9.3. Relevance of AI-based outputs

Security breaches can be created by an AI-driven application. Generally, any system that will be used to guide or advice on actions to be perform on network raise the same issue. For example, if an AI algorithm decides to change the filtering tables in a network it may compromise access control policies. Irrelevant results could be also produced. In the area of QoS, an AI system could allocate a bandwidth to a flow higher to the real link capacity. As shown from these two examples, an AI can produce decisions or values which are out of bounds of normal operations. To avoid such issues, safeguards can be added to discard or correct irrelevant outputs. Detecting such type of outputs can be also challenging in complex and distributed systems such as a network. Formal verification methods or testing techniques are helpful in that context.¶