EPDA: An Efficient and Privacy-preserving Dual Aggregation Scheme for Federated Learning in VANETs

Modern cars are no longer just machines for getting from point A to point B; they are rolling computers that constantly collect data about the road, the driver, and the surrounding traffic. This information is vital for making transportation safer and more efficient, but it also creates a significant privacy dilemma. Drivers do not want their location history or driving habits exposed, yet they need their vehicles to learn from the collective experience of the entire fleet. To solve this, researchers have developed a method called federated learning. Instead of sending raw data to a central server, each car trains a small computer model on its own local data and sends only the mathematical updates of that model to a coordinator. This keeps the private details on the vehicle, but it introduces a new problem: the updates themselves can still reveal sensitive information, and the open wireless channels used by cars are vulnerable to hackers who might pretend to be a car or tamper with the data.

A team of researchers from Jiangsu University has proposed a new system called EPDA to address these vulnerabilities specifically for vehicle networks. Their work focuses on creating a way for hundreds of cars to contribute to a shared learning model without revealing who they are or what they are driving, while also ensuring the process is fast enough to work in real-time traffic. The researchers designed a two-step process that handles both the verification of the cars' identities and the combination of their model updates. By using a specific type of digital signature that does not require heavy cryptographic calculations, they managed to keep the computational cost extremely low. In their tests using a standard dataset for traffic sign recognition, the system required only 0.03 seconds of processing time per car and transmitted just 148 bytes of data per client, even when scaling up to 350 vehicles. This efficiency suggests that such a system could realistically be deployed in busy city environments where speed and privacy are equally critical.

The core challenge in this field is balancing security with speed. Traditional methods for protecting data often rely on complex mathematical operations that take too long for the split-second decisions required in traffic. Other approaches might protect privacy but fail to hide the identity of the sender, leaving drivers vulnerable to tracking. The researchers behind EPDA tackled this by introducing a dual aggregation scheme. The first part of their system deals with authentication. In a typical scenario, a roadside unit acts as a semi-trusted collector, gathering updates from nearby cars. The researchers equipped this collector with a way to verify that all the incoming messages are genuine without having to check each one individually. They achieved this by grouping the digital signatures from multiple cars into a single, compact signature. This allows the roadside unit to confirm the validity of a whole batch of updates in one go, rather than slowing down to verify each vehicle one by one. This method avoids the need for heavy cryptographic tools that would otherwise bog down the system.

The second part of the system protects the actual content of the model updates. Even if a car is verified as legitimate, the data it sends could still be analyzed to infer private details. To prevent this, the researchers implemented a masking technique. Before a car sends its model update, it scrambles the data using a secret key shared with the central server. This scrambling ensures that the roadside unit, which is only semi-trusted, cannot see the raw model parameters. The central server, which holds the master key, can later unscramble the aggregated data to update the global model. This process guarantees that the intermediate collectors never see the sensitive information, while still allowing the final model to be reconstructed correctly. The system also ensures that the identity of the car remains hidden. Each vehicle uses a temporary, fake identity for its transmissions, making it impossible for an observer to link two different messages back to the same car. Only a trusted authority, kept separate from the data collectors, has the ability to reveal a real identity if a car behaves maliciously, ensuring that privacy does not come at the cost of total accountability.

To test how well this system works in practice, the researchers simulated a federated learning environment using a dataset of over 50,000 traffic sign images. They set up a scenario where anywhere from 50 to 350 virtual cars participated in the training process. The results showed that the time required for the cryptographic operations remained remarkably stable regardless of how many cars were involved. While the time needed for the cars to train their local models increased with the size of their data, the time spent on security tasks like masking and signing stayed consistently low, averaging less than 0.04 seconds per car. The time required for the central server to verify the batched signatures also grew linearly but remained very fast, taking only about 3.2 milliseconds to verify updates from 350 cars. In terms of data transmission, the system was highly efficient, requiring each car to send only 148 bytes of information per round. This is significantly less than other existing methods, which can require hundreds or even thousands of bytes for the same task.

The researchers compared their approach against several other state-of-the-art methods and found that their system was consistently faster and lighter. While some competing schemes relied on complex mathematical pairings that took nearly 10 milliseconds to verify, the EPDA system completed the same task in a fraction of that time. The communication overhead was also drastically reduced, with the total data size per car being less than a third of what some other methods required. These findings indicate that the proposed scheme successfully navigates the difficult trade-off between strong privacy guarantees and the high-speed demands of vehicular networks. The system proved capable of handling a large number of participants without degrading performance, suggesting it could scale effectively as the number of connected vehicles on the road continues to grow.

Ultimately, the work demonstrates that it is possible to build a secure, private, and efficient learning system for the automotive world without relying on the heavy computational machinery that has previously limited such applications. By combining a lightweight signature method with a secure masking technique, the researchers have created a framework where cars can learn together without exposing their secrets or slowing down the network. The simulations suggest that this approach is ready for real-world deployment, offering a practical path forward for intelligent transportation systems that respect user privacy while maintaining the high performance needed for safe driving. The study concludes that this dual aggregation method provides a viable solution for the future of connected vehicles, where data security and operational efficiency must go hand in hand.