Adaptive and Privacy-Preserving Federated Learning for Real-Time Smart Traffic Systems Using Hybrid Deep Models

• Abhishek P Nachankar,
• Nishat Afshan Ansari,
• Pooja Jain,

Vol. 19, No. 10, October 31, 2025
10.3837/tiis.2025.10.014, Download Paper (Free):

• Federated learning
• Smart Traffic
• privacy preservation
• Heterogeneous Data
• communication efficiency
• Scenarios

Abstract

There is a very urgent need for a privacy-preserving, communication-efficient, and adaptive learning framework for smart traffic systems to be deployed. Conventional centralized learning algorithms pose enormous threats to user privacy, bandwidth consumption, and computation, especially when dealing with miscellaneous and dynamic sources of data, such as traffic cameras, IoT sensors, and GPS devices & deployments. Less than ideal solutions to these issues are provided by existing federated learning (FL) models, which typically follow fixed schedules for updates, suffer from inefficiency in handling heterogeneous data, and provide limited assurance of privacy. To mitigate these shortcomings, we present a multi-method federated learning framework for real-time monitoring and management of traffic systems. The proposed framework weaves together five techniques, namely: (i) Privacy Adaptive Federated Contrastive Learning (PA-FCL), whereby contrastive learning is incorporated with adaptive privacy noise, such that sensitive data exposure is reduced while accuracy remains high; (ii) Hierarchical Adaptive Reinforcement Aggregation (HARA-FL), which utilizes multi-agent reinforcement learning to optimize update intervals and minimize communication burden; (iii) Graph Neural Network Driven Heterogeneous FL (GNN-FedHFL) to model the sensor data in a graph structure for learning improvements from heterogeneous data sources; (iv) Transformer-Enhanced Federated Vision Model (FedViT), leveraging Vision Transformers for obtaining robust visual feature extraction under adverse imaging conditions; and (v) Federated Adversarial Auto encoder with Differential Privacy (FAAE-DP) provision for extreme privacy through encoded and noise-added data representations. All the above methods collectively provide very significant benefits: 70% privacy risk reduction, 40% communication cost savings, 10 to 12% effectiveness increase on heterogeneous data, and 20% convergence delay reductions. Together, this suite of tools represents a jump forward for federated learning within the area of smart traffic systems and will facilitate the secure, efficient, and scalable realization in real-world settings.

Statistics

Cite this article