Intelligent Network Slicing for B5G ITS

The problem

One network, thousands of different needs

Autonomous cars, connected ambulances, streaming in the back seat — they all share the same network. But a collision alert cannot "wait in line" behind a YouTube video.

🚨

Vehicle safety

Cooperative braking, collision alerts — they require latency below 5ms and 99.999% reliability. A half-second delay can cost lives.

🗺️

Traffic efficiency

Real-time navigation, truck platooning, smart traffic lights — they need continuous, up-to-date data but can tolerate more latency.

🎬

Entertainment

4K streaming, in-vehicle videoconferencing — they need high bandwidth, but a 2-second buffer is not critical.

💡 Think of a highway with smart lanes

Imagine a highway where lanes reconfigure automatically based on traffic. During rush hour, the emergency lane expands. At night, more space goes to heavy cargo. Network slicing is exactly this — but in the digital infrastructure.

🚑

Safety Slice

Ultra-low latency, top priority

🚗

Efficiency Slice

Continuous data, high priority

📱

Generic Slice

High bandwidth, delay-tolerant

The research

Seven contributions, one narrative

The thesis progressively builds an ecosystem spanning from vehicular simulation foundations to the security of distributed artificial intelligence in critical transportation infrastructure.

🔐 Where networks, AI, and security converge

The research spans three intertwined axes. The first is B5G/6G mobile communications infrastructure, with network slicing and edge intelligence to meet the heterogeneous requirements of intelligent transportation systems (ITS). The second is artificial intelligence — supervised classification, federated learning, and hyperparameter optimization — which permeates nearly every contribution as the tool for making slice management adaptive and proactive. The third is security: introducing distributed AI into the infrastructure creates new attack surfaces, requiring threat modeling, adversarial attack assessment, and Byzantine-robust defenses. Systems operating in critical infrastructure cannot just be intelligent — they must be resilient.

Foundations: realistic vehicular network simulation

The starting point was understanding how to simulate vehicular networks faithfully — with signal propagation models, vehicle mobility, and interference. Published as a Springer book chapter, this work contributed to the understanding of vehicular network simulation fundamentals.

Book chapter — Springer, 2019

Detecting congestion with public data

Using GPS data from Rio de Janeiro's buses (31 million records), we selected two high-traffic regions — Lagoa Rodrigo de Freitas and Av. Ayrton Senna — and cross-referenced the data with congestion reports published by the city government on X (formerly Twitter). We trained 6 ML classifiers to detect congestion without dedicated sensors.

Accuracy > 80% across 6 classifiers

The network as an implicit sensor

We found that traffic changes alter the communication profile at edge servers. Instead of expensive sensors, the mobile communication network itself can serve as an implicit sensor of the vehicular context — using telemetry data already available in the infrastructure.

RF and DT with accuracy > 85%

A framework that prioritizes what matters

We built a framework based on SDN (Software-Defined Networking), a technology that enables centralized, dynamic network programming and reconfiguration. The framework organizes the network into dedicated slices for each application type (safety, efficiency, entertainment). When traffic conditions change, the control algorithms redistribute resources automatically. Evaluated with 158 vehicles in a realistic emulation scenario.

PDR of 85% vs. 50% baseline (+35 p.p.)

AIMS: classifying impact before the problem occurs

We created AIMS (Adaptive and Intelligent Management of Slicing) — a system that analyzes network telemetry in real time and classifies the impact on active slices into 4 levels: Adequate, Warning, Severe, and Critical. Instead of waiting for the network to fail, AIMS anticipates degradation and provides signals so that proactive reconfigurations can be triggered before the problem affects the user. Experimental validation focused on the impact classification function of the network slicing policies.

CatBoost: 94.8% accuracy (F1-Macro 0.948)

Distributed learning: each access point learns locally

Centralizing data from all RSUs (Road-Side Units — edge devices deployed along roads that connect vehicles to the network) is costly and compromises privacy. With federated learning, each RSU trains on its own data and shares only what it learned — not the raw data. Tested with 3 RSUs along an urban corridor under heterogeneous data scenarios.

Degradation < 0.5 p.p. vs. centralized

Information security: protecting distributed intelligence

Distributing the training process creates a new attack surface: a compromised node can execute model poisoning by combining targeted label-flipping with gradient scaling. In practice, the attacker manipulates the classifier so that critical situations are classified as normal — suppressing all response actions for safety-critical applications.

We diagnosed the vulnerability across 144 experimental configurations (3 neural architectures × 6 aggregation strategies × multiple attack factors) and evaluated Byzantine-robust aggregation defenses (Krum and Trimmed Mean). The result: without defense, the misclassification rate reaches 100%; with Krum, it drops to zero — demonstrating that aggregation-layer security is a requirement for safe deployment of distributed AI in critical infrastructure.

No defense: 100% deception | With Krum: 0%

Publications

Scientific output

4 published papers, 2 submitted, and 1 in preparation — in IEEE and Springer venues. Buttons on each publication link to the original paper. Full profile on ORCID and Google Scholar.

T-ITS

An Application-Driven Framework for Intelligent Transportation Systems Using 5G Network Slicing

IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 8, 2021 — Saraiva, Campos, Fontes, Rothenberg, Sorour & Valaee

DOI Code

IWCMC

An Approach Using Machine Learning and Public Data to Detect Traffic Jams

IEEE International Wireless Communications and Mobile Computing Conference (IWCMC), 2021 — Saraiva & Campos

DOI

VTC

Machine Learning Applied to Network Traffic for Vehicular Applications to Predict the Vehicle Context

IEEE 94th Vehicular Technology Conference (VTC2021-Fall), 2021 — Saraiva, Girardi & Campos

IEEE Xplore

Book

Realistic Vehicular Networks Simulations with Propagation Models and Mobility Traces

Springer — Smart and Digital Cities: From Computational Intelligence to Applied Social Sciences, 2019 — Saraiva & Campos

DOI

Submitted

AIMS: Adaptive and Intelligent Management of Slicing

IEEE International Symposium on Computers and Communications (ISCC), 2026 — Saraiva & Campos

Submitted

Federated Learning-Based Scalable Impact Classification for ITS Network Slicing

IEEE Vehicular Technology Conference (VTC2026-Spring), 2026 — Saraiva & Campos

In review

Poisoning Attacks and Byzantine-Robust Defenses in Federated Impact Classification for ITS Network Slicing

Manuscript under review for submission, 2026 — Saraiva & Campos — Security assessment of distributed AI across 144 experimental configurations

Open source & reproducibility

All experiments are reproducible. Code available on GitHub and CodeOcean.

framework_its_sdn

SDN framework with 5G slicing for ITS applications (Ch. 5)

AIMS

Centralized and federated impact classification + security assessment — RF, CatBoost, DNN, LSTM, GRU + Krum, TM (Ch. 6–8)

Frequently asked questions

Common questions

What is "network slicing"? +

It is the ability to divide a single physical network into multiple isolated virtual networks, each optimized for a specific use. Think of dedicated highway lanes: one for ambulances (ultra-low latency), another for trucks (continuous data), another for passenger cars (general use). Each "slice" has its own guaranteed resources without interfering with the others.

Does this work on real 5G networks? +

Experiments were conducted in a realistic emulation environment (Mininet-WiFi + SUMO), which faithfully reproduces wireless network behavior and vehicular mobility. The algorithms operate at the control layer (SDN), which is independent of the radio access technology. Validation on real 5G testbeds is the natural next step.

What is "federated learning"? +

It is a way to train AI models without centralizing data. Each roadside access point (RSU — Road-Side Unit) trains on its own data and shares only what it learned (model parameters), not the data itself. It is like hospitals collaborating on a medical study by sharing conclusions but never sharing patient records.

Why worry about security? +

When we distribute training across multiple nodes, we create a new attack surface: a compromised node can "poison" the global model, causing the system to classify critical situations as normal. In a vehicular context, this means the system could fail to react when a safety slice is degrading — with potentially fatal consequences. The research demonstrated that defenses like Krum reduce this deception rate to zero.

Where and when was the research conducted? +

The PhD was conducted at the Graduate Program in Informatics (PPGI) at UNIRIO (Federal University of the State of Rio de Janeiro), under the supervision of Prof. Dr. Carlos Alberto Vieira Campos. The defense took place in March 2026, with approval by a committee composed of researchers from UNIRIO, UNICAMP, and UFRJ.

Can I reproduce the experiments? +

Yes. All code is available on GitHub and CodeOcean. Datasets are open. Experiments use deterministic seeds for reproducibility. The computational environment is documented in the thesis tables.

How do attacks and defenses work in distributed AI? +

By distributing training via federated learning, we create an attack surface that goes beyond traditional network threats. The attack investigated in the thesis is model poisoning: a malicious participant flips its data labels (turning "critical" into "normal") and amplifies the weights sent to the server to dominate the aggregation. The result is that the system stops detecting risk situations — in transportation infrastructure, this could prevent the network from reacting when an autonomous vehicle needs it.

The evaluated defenses are Byzantine-robust aggregation strategies (Krum and Trimmed Mean), which filter malicious updates before aggregation — functioning as a firewall for AI models. The thesis quantifies the operational risk across 144 configurations and demonstrates that, with Krum, the deception rate drops to zero. The logic is the same as a pentest: diagnose the vulnerability, measure the impact, and validate the defense.

How to securely manage 5G/6G networks to optimize application requirements in intelligent transportation systems?