Federated Intrusion Detection Framework for 5G Networks Aligned with ETSI NFV MANO

Researchers at the University of Portsmouth have unveiled V5GIDS, an artificial intelligence-driven intrusion detection framework designed to work natively inside the management and orchestration machinery of modern 5G networks. The framework, described in a new paper in the Journal of Network and Systems Management, addresses a persistent gap in network security research: many intrusion detection systems perform impressively in laboratory benchmarks but fail to fit the operational realities of virtualized, service-based 5G infrastructures, where lifecycle automation, orchestration-aware deployment and telemetry efficiency are not optional extras but core requirements.

The team, led by Sofian Ben Khalifa together with Rahim Taheri and Ivan Jordanov, argues that the security community has been optimizing metrics while neglecting deployment relevance. Since the first commercial 5G network was switched on in 2018, operators have moved away from monolithic hardware appliances and toward software-defined, virtualized network functions. This transformation brings enormous flexibility, but it also enlarges the attack surface. Virtualization, service-based control, edge distribution and highly dynamic traffic conditions collectively introduce management and security challenges that a standalone machine learning model, trained offline on static packet captures, was never designed to meet.

V5GIDS is built from the ground up around the ETSI NFV MANO ecosystem, the European Telecommunications Standards Institute’s framework for managing and orchestrating network functions virtualization. Rather than treating an intrusion detector as an external bolt-on, the researchers model the detector itself as a Virtualized Network Function, or VNF, deployable and lifecycled by the orchestrator like any other network component. Monitoring, analysis, policy propagation and model management are then organized within a closed-loop security orchestration process, aligned with the emerging Zero-Touch Service Management principles that aim to let networks configure, heal and defend themselves with minimal human intervention.

The technical heart of the framework is a two-stage detection pipeline. The first stage uses a convolutional autoencoder, a neural architecture that learns to compress and reconstruct representations of network flow features. Stacked convolutional autoencoders of this kind, first popularized for hierarchical feature extraction more than a decade ago, excel at distilling high-dimensional traffic descriptors into compact latent representations. In V5GIDS, the autoencoder stage transforms raw network-flow data into learned embeddings, suppressing noise and dimensionality before classification. The second stage applies a random forest, an ensemble of decision trees famously introduced by Leo Breiman in 2001, which performs the final judgment: benign or malicious. The combination is pragmatic. Autoencoders handle representation learning on unlabelled and heterogeneous traffic; random forests provide fast, interpretable classification with low inference overhead, an important property for resource-constrained edge environments.

Crucially, the framework avoids centralizing raw traffic data. Network operators are understandably reluctant to ship enormous volumes of flow records to a central server, both because of the bandwidth cost on the control plane and because of privacy and regulatory concerns. Instead, V5GIDS adopts federated learning, a distributed training strategy in which each participating site trains a local model on its own data and shares only model updates. The aggregated global detector learns from the collective experience of the network without any single node ever seeing another node’s raw flows. In a further twist aimed at keeping telemetry exchange minimal, the framework trains the detector using only 2 to 3 percent of the available network-flow data, a deliberate demonstration that meaningful security intelligence can be extracted from a thin, carefully chosen slice of network telemetry rather than exhaustive data collection.

The researchers validated their design with a working prototype spanning five federated clients, simulating the distributed character of a real 5G deployment in which detection functions would run across network edges, cores and operator domains. For evaluation, they turned to two widely used benchmark collections. The first is CIC-IDS2017, a canonical intrusion detection dataset from the Canadian Institute for Cybersecurity that characterizes realistic attack scenarios such as brute force, denial-of-service and web attacks alongside benign traffic. The second is the newer CIC-BCCC-NRC-TabularAttacks-2024, a more recent dataset of tabular attack representations reflecting contemporary threat patterns. Testing against both an established baseline and a cutting-edge collection gives a fuller picture of how the pipeline behaves across different generations of threat data.

The results show that V5GIDS achieves strong detection performance in the evaluated settings, while the more significant contribution may lie elsewhere. By aligning the detector with orchestration standards, the framework improves deployment relevance in ways that raw accuracy figures do not capture. The closed-loop design means that when a new model version is ready, it can be propagated through the orchestration layer using standardized interfaces, consistent with the YANG-based descriptors and RESTful protocols specified in recent ETSI NFV releases. Policy changes, model retraining and detector scaling become orchestration operations rather than bespoke engineering projects, reducing the operational friction that has kept many academic intrusion detection systems out of production networks.

The resource-aware, distributed character of the system is equally consequential for 5G operators. Edge deployments face tight computational budgets, and shipping large telemetry streams to centralized security platforms consumes both bandwidth and processing capacity that could serve customers. By keeping training local, exchanging only lightweight model updates, and demonstrating effective learning from a 2 to 3 percent data sample, V5GIDS charts a path toward security functions that live comfortably within the constraints of the environments they protect. This matters as network slicing and edge computing multiply the number of distinct environments requiring protection, each with its own traffic profile and threat exposure.

The work also speaks to a broader tension in 5G security architecture. Standards bodies including 3GPP, ETSI and the GSM Association have documented how the 5G system’s service-based architecture, network slicing and virtualization create novel security considerations, and industry surveys consistently flag security orchestration as an unresolved challenge for next-generation networks. Prior academic efforts have produced self-adaptive deep learning anomaly detectors for 5G and intrusion detection systems with explainability mechanisms for the 5G core, but these have often remained weakly connected to the management plane. V5GIDS deliberately reverses that emphasis: the machine learning is competent, but the architectural integration is the point.

The researchers have released the source code and implementation on a public repository, allowing other teams to reproduce the five-client federated prototype and to test the two-stage pipeline on their own traffic data. That openness matters for a field where reproducibility is frequently undermined by proprietary datasets and unshareable prototypes. Combined with the use of publicly available benchmark datasets and standards-aligned architecture, the release lowers the barrier for operators and researchers who want to experiment with orchestration-native security in testbed environments before committing to production deployment.

There are, of course, limits to what the current evaluation establishes. Federated learning introduces its own vulnerabilities, including the possibility of poisoned model updates from compromised clients, and the paper’s datasets, while well regarded, cannot capture every adversarial behavior a determined attacker might mount against a live 5G network. The 2 to 3 percent training regime, while impressive, will require validation across more diverse operator environments and traffic mixes. And the gap between a five-client prototype and a multi-domain, multi-vendor production network remains substantial, involving questions of trust, governance and cross-operator data protection that no single framework can resolve alone.

Nevertheless, V5GIDS offers a concrete template for what deployment-ready intrusion detection could look like as networks grow more virtualized and more autonomous. Its core insight is architectural as much as computational: a security function earns its place in a 5G network not only by detecting attacks accurately, but by fitting the orchestration, automation and telemetry constraints that define how such networks are actually managed. As the industry pursues zero-touch operations, security functions that cannot participate in the closed loop risk becoming dead weight. This framework shows what it takes to join the loop, and in doing so narrows the stubborn distance between laboratory promise and operational practice in 5G network defense.