🤖 machine learning

Effective and Efficient Cross-City Traffic Knowledge Transfer: A Privacy-Preserving Perspective

This paper proposes FedTT, a novel privacy-preserving federated learning framework that enhances cross-city traffic knowledge transfer by integrating traffic view imputation, domain adaptation, and secret aggregation to overcome challenges related to data quality, distribution discrepancies, and privacy leakage.

Original authors: Zhihao Zeng, Ziquan Fang, Yuting Huang, Lu Chen, Yunjun Gao

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhihao Zeng, Ziquan Fang, Yuting Huang, Lu Chen, Yunjun Gao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a new driver how to navigate a specific city, but that city has very few traffic cameras and not enough data to learn from. Meanwhile, three other big cities have mountains of traffic data, but they are legally forbidden from sharing their raw data (like license plate numbers or specific routes) because of privacy laws.

This is the problem the paper FedTT tries to solve. It's a new system that lets these "data-rich" cities teach the "data-poor" city how to predict traffic without ever handing over their private files.

Here is how FedTT works, broken down into simple concepts and analogies:

The Four Big Problems

Before building their solution, the authors identified four major hurdles in current methods:

The Privacy Leak: Even if you don't share the raw data, sharing the "math" used to learn from it (gradients) can sometimes let hackers reverse-engineer the original data. It's like sharing a recipe that accidentally reveals the secret ingredient.
The "Apples to Oranges" Problem: Traffic in San Francisco looks very different from traffic in Nashville. Trying to teach a San Francisco driver how to drive in Nashville using San Francisco's rules doesn't work well. The data distributions are too different.
The Broken Sensor Problem: In the real world, sensors break or go offline. If a city's data has holes in it, the learning model gets confused and makes mistakes.
The Slow & Expensive Problem: Current methods take a long time to train and require massive amounts of data transfer, making them too slow and expensive for real-world use.

The FedTT Solution: A Four-Step Kitchen

The authors created a framework called FedTT (Federated Traffic Knowledge Transfer). Think of it as a master chef (the target city) trying to learn a new cuisine from three other chefs (source cities) who won't let the master chef into their kitchens.

Here are the four "tools" FedTT uses to solve the problems:

1. The "Data Repair Kit" (Traffic View Imputation - TVI)

The Problem: Sensors break, leaving gaps in the data.
The Analogy: Imagine a puzzle with missing pieces. Instead of giving up, FedTT uses a "repair kit." It looks at the surrounding pieces (spatial neighbors) and the pieces from the previous and next moments in time (temporal neighbors) to guess what the missing piece should look like.
The Result: It fills in the holes in the data so the model learns from a complete picture, not a broken one.

2. The "Universal Translator" (Traffic Domain Adapter - TDA)

The Problem: San Francisco traffic (hilly, winding) is different from San Bernardino traffic (flat, grid-like). You can't just copy-paste knowledge.
The Analogy: Think of this as a translator that doesn't just translate words, but changes the accent and dialect. FedTT takes the traffic patterns from the source cities and "translates" them so they sound exactly like the traffic patterns of the target city. It uses a "Generator" (the translator) and a "Discriminator" (the critic) in a game of cat-and-mouse to ensure the translated data looks 100% authentic to the target city.
The Result: The target city learns from the source cities, but the data feels like it came from its own streets.

3. The "Secret Handshake" (Traffic Secret Transmission - TST)

The Problem: How do you share the "translated" data without revealing the original secrets?
The Analogy: Imagine the source cities want to send a secret message to the target city. Instead of sending the message directly, they each add a random "mask" to their data. They send these masked versions to a central server. The server adds them all up. Because the masks cancel each other out mathematically, the server gets the average of all the data, but it cannot see any single city's individual data.
The Result: The target city gets the collective wisdom of the group, but no single city's private data is ever exposed. It's secure, but unlike heavy encryption methods, it's fast and doesn't slow things down.

4. The "Parallel Assembly Line" (Federated Parallel Training - FPT)

The Problem: Doing everything one step at a time is too slow.
The Analogy: Instead of a single worker building a car from start to finish, FedTT sets up an assembly line where different teams work on different parts of the car at the same time. While one team is fixing the engine (training the generator), another is painting the body (training the discriminator). They freeze certain parts of the data so they don't have to wait for each other.
The Result: The whole system trains much faster and uses less bandwidth.

The Results

The authors tested this system on real traffic data from four cities: San Francisco, San Bernardino, Nashville, and Hong Kong.

Accuracy: FedTT made much better traffic predictions than 18 other existing methods. It reduced prediction errors significantly (by about 5% to 22%).
Speed: It was dramatically faster, running 10 to 40 times faster than other methods because of the parallel training and efficient data handling.
Privacy: It successfully protected the data without needing heavy, slow encryption.

Summary

FedTT is like a smart, secure, and fast way for cities to share traffic wisdom. It fixes broken data, translates different traffic styles into a common language, keeps everyone's secrets safe using a clever math trick, and does it all at the speed of light. This allows new or data-poor cities to learn from experienced ones without breaking privacy laws.

Technical Summary: FedTT – Effective and Efficient Cross-City Traffic Knowledge Transfer

Problem Definition
Traffic prediction (TP) is critical for urban management but faces significant challenges in emerging cities where data scarcity leads to overfitting and poor model accuracy. While Transfer Learning (TL) offers a solution by leveraging data-rich source cities, existing Federated Traffic Knowledge Transfer (FTT) approaches suffer from four unresolved challenges:

Privacy Leakage: Although federated learning avoids raw data sharing, uploading gradients or model parameters still risks inference attacks that can reconstruct sensitive traffic data.
Distribution Discrepancy: Significant differences in traffic patterns (flow, speed, occupancy) between source and target cities reduce the effectiveness of knowledge transfer.
Low Data Quality: Real-world sensor data often contains missing values due to failures or updates, which existing methods handle poorly (e.g., by defaulting to zero), degrading model performance.
Inefficiency: Current FTT methods rely on serial two-stage processes and complex models, resulting in high computational and communication overheads (often 40x slower than non-transfer baselines).

Methodology: The FedTT Framework
The authors propose FedTT, a privacy-preserving federated framework that transforms source city traffic data into the target city's domain before training, rather than simply aggregating model parameters. The framework consists of four key modules:

Traffic View Imputation (TVI):
- Goal: Address low data quality and missing sensor readings.
- Mechanism: Before federated training, TVI completes missing data by capturing spatio-temporal dependencies. It employs a Spatial View Extension using Graph Attention Networks (GAT) and Dijkstra's algorithm to infer missing sensor values based on road network topology. It then uses a Temporal View Enhancement (leveraging the DyHSL model) to predict missing values based on historical and future temporal sequences.
- Outcome: Generates a complete traffic view ( $\hat{X}_t$ ) for all sensors, ensuring robust training even with incomplete data.
Traffic Domain Adapter (TDA):
- Goal: Mitigate cross-city distribution discrepancies.
- Mechanism: TDA uses a Generative Adversarial Network (GAN) architecture to uniformly transform data from the source domain to the target domain.
  - Transformation: A generator ( $\theta_{Gen}$ ) maps source data to the target domain using transformation matrices derived from road network structures and traffic domain prototypes (averaged traffic data).
  - Alignment & Classification: A discriminator ( $\theta_{Dis}$ ) ensures the transformed data aligns with the target domain's distribution. The generator is trained to fool the discriminator, effectively aligning the source data distribution with the target city's specific traffic patterns.
Traffic Secret Transmission (TST):
- Goal: Protect privacy during aggregation without sacrificing efficiency (avoiding heavy Homomorphic Encryption or accuracy-reducing Differential Privacy).
- Mechanism: Instead of uploading raw transformed data or gradients, clients transmit masked data.
  - In the first round, clients encrypt and send initial data.
  - In subsequent rounds, clients compute a mask based on the difference between the current transformed data and the previous aggregated data.
  - The server aggregates these masks to reconstruct the new aggregated data.
- Outcome: The server only accesses the aggregated result, preventing the inference of individual client data while maintaining high training efficiency.
Federated Parallel Training (FPT):
- Goal: Reduce communication overhead and improve training speed.
- Mechanism: FPT employs Split Learning and Parallel Optimization. It decomposes the training process, allowing clients and the server to train specific modules (e.g., generator, discriminator, traffic model) simultaneously. It freezes intermediate results (e.g., classifier outputs, aggregated data) to break data dependencies, enabling parallel execution and significantly reducing data transmission.

Key Contributions

Framework Design: Proposes FedTT, a novel framework that shifts the paradigm from model aggregation to data domain transformation within a federated setting.
Robustness: Introduces the TVI module to handle missing data via spatio-temporal view extension, improving robustness in real-world scenarios.
Effectiveness: Develops the TDA module to explicitly align heterogeneous source-target distributions, enhancing transfer effectiveness.
Privacy & Efficiency: Designs the TST module for lightweight, secure aggregation and the FPT module for parallel training, addressing both privacy leakage risks and computational inefficiency.
Empirical Validation: Demonstrates state-of-the-art performance across multiple datasets and prediction tasks.

Experimental Results
The authors evaluated FedTT on four real-world datasets (PeMSD4, PeMSD8, FT-AED, HK-Traffic) covering San Francisco, San Bernardino, Nashville, and Hong Kong.

Performance: FedTT outperformed 18 state-of-the-art baselines (including T-ISTGNN, pFedCTP, and various MTT/STT methods) across traffic flow, speed, and occupancy prediction tasks.
- Accuracy: Achieved a reduction in Mean Absolute Error (MAE) ranging from 5.43% to 22.78% and RMSE reductions from 7.25% to 50.79%.
- Scalability: Maintained superior performance even when the target city's training data was reduced to 5%.
- Long-term Prediction: Successfully handled long-term predictions (60 minutes), showing gains of 6.86% to 17.30% over baselines.
Efficiency: FedTT reduced communication overhead by 90% and running time by 1 to 2 orders of magnitude compared to existing FTT methods.
Ablation Studies: Confirmed that each module (TVI, TDA, TST, FPT) contributes significantly to the overall performance, with TDA and TVI being critical for accuracy and TST/FPT for efficiency and privacy.

Significance and Claims
The paper claims that FedTT provides a comprehensive solution to the "four challenges" of existing FTT methods. By transforming the data domain rather than just aggregating models, FedTT effectively bridges the gap between data-rich and data-scarce cities without compromising privacy. The authors assert that their lightweight secret aggregation and parallel training strategies make the framework viable for real-world deployment where data privacy regulations (like GDPR) and computational resources are limiting factors. The work establishes a new baseline for privacy-preserving, efficient, and robust cross-city traffic knowledge transfer.