St. Petersburg Researchers Boost Accuracy of a Safer Way to Train Smart City AI

Russia’s Secure AI Push for City Systems

A team at Saint Petersburg Electrotechnical University “LETI” has reported progress in secure machine learning for city-scale digital systems, an area that matters for the resilience of services such as monitoring, public safety and transportation. The work builds on federated learning, which improves model quality while keeping data private by avoiding centralization.

The results underscore local scientific advances in Russia’s privacy-preserving learning and strengthen the country’s position in building safer AI systems for smart city use cases. The algorithm is positioned to improve the effectiveness of city analytics and monitoring while reducing data leakage risk. The modified version, MFedBN, could also be of interest to the international privacy-preserving machine learning community.

Federated learning makes it possible to train models without sending local datasets to a central server while still reaching high accuracy in security and transportation analytics. MFedBN’s advantage is greater stability and adaptability when data are heterogeneous across participants.

Export Pathways and Domestic Use Cases

Privacy-preserving by design, federated learning has become a high-demand direction in AI internationally, with applications spanning healthcare, finance and telecommunications. Algorithms in the MFedBN family can open export opportunities for Russian AI technologies in commercial and research projects, including IoT, smart transportation and cybersecurity.

Inside Russia, near-term pathways include integrating the approach into city infrastructure management systems, deploying it in distributed and privacy-focused cybersecurity setups (IDS/IPS), and extending it to IoT devices where centralized training is difficult.

The viability and superiority of the proposed algorithm over baseline alternatives were demonstrated through a series of experiments on two fundamentally different tasks – monitoring commercial transport behavior and ensuring network security. Specifically, when tested on data from truck sensors, the algorithm achieved 85% classification accuracy, and in detecting cyberthreats and network intrusions it reached 99.98%

Ivan Kholod

research project lead and vice rector for digital transformation at Saint Petersburg Electrotechnical University “LETI”

With interest in privacy-preserving machine learning rising across research communities, federated learning has room to expand into interconnected networks such as smart transportation, healthcare delivery and energy. But scaling it successfully depends on solving practical challenges, including aligning standards and training protocols among participants and recognizing that a final model’s performance depends directly on how local data are distributed and how the network is designed.

How Federated Learning Got Here

In 2021, researchers introduced FedBN: “Federated Learning on Non-IID Features via Local Batch Normalization.”

In 2023, FedWon was announced as a new approach to federated learning (FL) without normalization. The work targeted multi-domain FL, where clients’ data come from different domains with different feature distributions rather than label distributions. Results showed FedWon outperforming FedAvg and FedBN across experimental settings.

A 2024 review of federated learning for intrusion detection in IoT (Leonardo et al.) reinforced FL’s promise for network security in distributed IoT ecosystems. FL allows attack-detection models to be trained directly on devices without transferring raw data, supporting confidentiality and improving robustness against new attack types.

Federated learning (FL) is gaining traction as a core AI approach. It enables training on distributed datasets without consolidating them into a single center, which reduces the risk of leaks and unauthorized access.

Federated learning is advancing with support from both research and business communities. The field has evolved from specialized solutions built for particular data challenges (such as FedBN) toward modern aggregation methods and broader deployment across smart-device environments (IoT).

Toward Comprehensive Data Protection for City Analytics

Russian researchers have made a notable contribution to federated learning by improving model accuracy and stability – a meaningful step for city systems that operate on distributed data. MFedBN’s performance demonstrates strong potential for AI deployments in security and network analytics, making them appealing to practical city agencies and IT companies.

Forecasts suggest federated learning could become an industry standard for privacy-preserving machine learning in critical services by 2030, including security, healthcare and telecommunications, where data protection is paramount. Stronger Russian research positions could raise the export potential of domestic AI solutions for markets with strict privacy expectations. Integrations with other advanced techniques – such as homomorphic encryption and differential privacy – are also expected.