Balancing Data Privacy and Utility in Trajectory Data: A Collaborative Adversarial Learning Approach

Imagine you are a detective tracking a suspect through a city. The suspect avoids using their real name, but their patterns of movement—home, work, favorite café—give them away. Even without an official ID, their unique trajectory reveals who they are.

This is exactly how Trajectory-User Linking (TUL) attacks work. Even when personal identifiers are removed, attackers use machine learning to re-identify users from their trajectory patterns.

While anonymization techniques help, they often degrade data utility, making it hard for businesses and researchers to extract meaningful insights. How can we protect user privacy without ruining data usability? Enter BPUCAL—a collaborative adversarial learning model that misleads TUL attacks while keeping trajectory data useful.

Challenges in Trajectory Privacy Protection

Existing privacy methods struggle with two major issues:

1. The Privacy-Utility Trade-off

   • Techniques like differential privacy and k-anonymity introduce noise or generalization, but this distorts data, reducing its value for traffic prediction, POI recommendations, and urban planning.

2. TUL Models are Too Powerful

   • Attackers use deep learning models to compare anonymous trajectories with past mobility data to re-identify users. Simple anonymization isn’t enough.

BPUCAL solves this by intelligently modifying only the most crucial trajectory points, balancing privacy protection and data usability.

BPUCAL: How It Works

BPUCAL is built on a three-module adversarial learning system:

1. Trajectory Embedding Module

   • Converts raw trajectory data into a low-dimensional vector representation using LSTMs.
   • Captures spatiotemporal dependencies while reducing computational complexity.

2. Collaborative Adversarial Learning Module

   • Contains three key components:
     • Perturbation Generator (G): Identifies and perturbs the most critical trajectory points.
     • Discriminator (D): Ensures the perturbed trajectory remains realistic.
     • TUL Model: Helps train the generator by pinpointing high-risk trajectory segments.
   • Uses adversarial training to create realistic yet misleading trajectory variations.

3. Perturbed Trajectory Generation Module

   • Applies a softmax-based transformation to reconstruct trajectories while maintaining utility.

This method targets only high-risk trajectory points rather than blindly perturbing all data, preserving data quality while ensuring privacy.

Code Example: How BPUCAL Perturbs a Trajectory

Below is a Python snippet demonstrating how BPUCAL perturbs high-risk trajectory points:

import numpy as np

def perturb_trajectory(trajectory, critical_indices, epsilon=0.1):
    """Applies small perturbations to critical trajectory points."""
    perturbed_trajectory = trajectory.copy()
    for idx in critical_indices:
        perturbed_trajectory[idx] += np.random.uniform(-epsilon, epsilon, size=trajectory[idx].shape)
    return perturbed_trajectory

# Example trajectory (simplified)
original_trajectory = np.array([[35.6895, 139.6917], [34.0522, -118.2437], [40.7128, -74.0060]])  # Lat-Long pairs
critical_points = [1]  # Assume the 2nd point is crucial for identification
perturbed_trajectory = perturb_trajectory(original_trajectory, critical_points)

print("Original:", original_trajectory)
print("Perturbed:", perturbed_trajectory)

This function identifies critical trajectory points and applies small, targeted perturbations, ensuring minimal distortion while misleading attackers.

Experimental Validation: Does BPUCAL Work?

BPUCAL was tested on two real-world datasets:

• Foursquare (50,414 trajectories, 9,458 POIs)
• Weeplaces (101,771 trajectories, 22,140 POIs)

The results showed: ✅ 50% drop in TUL model accuracy while maintaining 90% of data utility. ✅ Outperformed random perturbation and heuristic-based methods. ✅ Maintained high accuracy in downstream tasks like POI recommendations.

Takeaways: Why BPUCAL Matters

🔹 Stronger Privacy Protection: TUL models can no longer easily re-identify users.

🔹 Minimal Data Distortion: Unlike naive perturbation, BPUCAL keeps data useful.

🔹 Scalable for Real-World Use: Works well for urban analytics, AI-driven mobility models, and location-based services.

What’s Next?

• Adapting BPUCAL for real-time, dynamic trajectory perturbation.
• Exploring federated learning to reduce privacy risks further.
• Combining adversarial learning with encryption-based privacy protection.

By protecting user trajectories without ruining data quality, BPUCAL offers a future-proof solution for privacy-preserving mobility analytics. 🚀

Reference

Lun, Y., Miao, H., Shen, J., Wang, R., Wang, X., & Wang, S. (2024). Resisting tul attack: balancing data privacy and utility on trajectory via collaborative adversarial learning. GeoInformatica, 28(3), 381-401.