Russian Researchers Train Neural Networks for Smart Cities
Developed by researchers in St. Petersburg, a new neural network training method allows models to learn safely from fragmented data while achieving near-perfect accuracy in cybersecurity-related tasks.
Training as a Federation
Researchers at Saint Petersburg Electrotechnical University “LETI” have refined a neural network training method used in smart city systems. They modified a federated learning algorithm, an approach in which a model is trained directly on data stored on local computing devices rather than on a central server or in the cloud. This data-local approach helps preserve confidentiality and reduces the risk of data leaks.
No One-Size-Fits-All Averaging
In conventional federated learning, the process is relatively straightforward. Each smartphone, camera, or sensor trains its own local version of a neural network using its data and then sends the results to a server. The server aggregates and averages these updates. When some devices observe one type of pattern and others see something entirely different, simple averaging inevitably distorts the picture. The model becomes confused and starts producing unreliable results.
LETI researchers addressed this limitation by redesigning the aggregation logic. Instead of treating all updates equally, their algorithm evaluates where the data comes from and adjusts the weight of each device’s contribution. Inputs that deviate sharply from the rest are downweighted, while stable and consistent data has greater influence. This adaptive averaging produces results that more accurately reflect real-world conditions.
Stable Learning Without Spikes
As a result, training becomes far more stable. The model no longer jumps erratically between conflicting data sources or struggles to determine which inputs are more reliable. This stability is particularly critical for smart city systems, where data flows in from cameras, sensors, and networked devices operating under vastly different conditions.
The researchers also developed a method for generating test datasets that allow the algorithm to be evaluated in scenarios that closely mirror real-world environments.
According to project lead Ivan Kholod, vice rector for digital transformation at ETU “LETI,” the updated algorithm delivered striking results. In detecting network attacks and cyber threats, the model achieved an accuracy of 99.98 percent. In analyzing traffic flow behavior, accuracy reached 85 percent — significantly outperforming baseline methods.
Developing and training such algorithms is crucial for urban infrastructure management. In smart cities, neural networks process massive data streams to control transportation systems, energy grids, security services, and environmental monitoring. Ensuring data privacy while maintaining high training and operational accuracy makes it possible to build more reliable services without exposing personal or sensitive information.
Beyond smart cities, the new algorithm could also be applied in other fields that require secure learning on distributed data, including healthcare and industrial analytics.
