Page 374 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

surges—in one slice can affect others. While deploying separate anomaly detection models
per slice is possible, their effectiveness is limited. These models rely only on slice-local data,
which may miss anomalies that appear similarly across multiple slices.

Data sharing between slices is often restricted due to privacy concerns, and exchanging raw
data can lead to redundant training processes. To address these challenges, we propose
a federated learning-based anomaly detection framework that enables cross-slice learning
without compromising data privacy.

This approach leverages the Network Data Analytics Function (NWDAF) as defined in Section
5 of 3GPP TS 23.288 [3], using standardized interfaces for secure and efficient exchange of
model updates. Key features such as message integrity and compression ensure low-overhead
and trustworthy communication. When an anomaly is detected in one slice, its signature is
generalized through federated learning and shared across all participating slices as illustrated
in Figure TBD. This enables real-time anomaly detection and mitigation in other slices, even
before similar issues occur for which the local model isn't trained to detect.

The framework maintains strict data isolation, reduces detection latency, and strengthens overall
network reliability. It aligns with both 3GPP architectural guidelines and ITU-T AI integration
frameworks, emphasizing privacy-preserving, distributed intelligence in telecom networks.

Use Case Status: Implementation in progress.

Partners : No partners

2�2 Benefits of the use case

This solution enhances 5G network security and resilience through AI-driven anomaly detection,
contributing to the development of reliable and sustainable communication infrastructure. By
proactively identifying and mitigating disruptions, it ensures the consistent performance of
critical network systems.

The system also supports smart city applications by improving overall network reliability and
security. This reduces service disruptions and enables seamless connectivity for Internet of
Things (IoT) services, autonomous systems, and digital public infrastructure.

In addition, the federated learning-based approach strengthens cybersecurity and data
privacy by preventing fraud and cyber threats. It promotes trust and transparency in
digital communications while safeguarding sensitive information in critical communication
environments.

2�3 Future Work

• Planned real-world deployment: The framework can be validated using open-source
emulators (e.g., srsRAN+srsUE, Free5GC). This deployment will assess the operational
robustness of the federated learning system under realistic traffic loads and network
slicing dynamics.
• Improvements to adversarial and distributed training: Future work will incorporate
robust federated learning methods—such as differential privacy, secure aggregation, and
Byzantine-resilient averaging—to defend against poisoning attacks and maintain model
integrity in untrusted or distributed environments.

338