TL;DR
This paper introduces PassiFi, a novel indoor positioning system that leverages passive Wi-Fi Time Difference of Arrival (TDoA) measurements and deep learning to achieve sub-meter accuracy while prioritizing user privacy. By avoiding active signal transmissions, PassiFi mitigates privacy risks and spoofing attacks inherent in existing Wi-Fi RTT-based localization methods.
What Problem Does This Paper Address?
Accurate indoor localization is crucial for numerous applications, from navigation to security. However, existing Wi-Fi-based localization systems, particularly those using Wi-Fi Fine Timing Measurement (FTM) for Round Trip Time (RTT) estimation, often rely on active signal transmissions. This "active" nature means the user's device must send signals, making it detectable and vulnerable to adversaries who can track movements or even manipulate measurements. This poses a significant privacy concern. Furthermore, indoor environments are challenging due to multipath propagation and non-line-of-sight (NLOS) conditions, which degrade the accuracy of many localization techniques. The paper aims to bridge this gap by developing a system that is both highly accurate and privacy-preserving.
Methods
PassiFi tackles the problem by employing a passive Wi-Fi TDoA approach. Instead of actively sending signals, the user's device passively listens to Wi-Fi frames exchanged between access points (APs). It then calculates the Time Difference of Arrival (TDoA) of these signals. This passive collection of TDoA values, combined with a fingerprinting technique, forms the core of the system.
Here's a breakdown of the key methods:
- Passive TDoA Measurement: The system utilizes the IEEE 802.11mc standard's FTM protocol but in a passive manner. The user's device (the "passive station") records the arrival times of FTM frames and their corresponding acknowledgments (ACKs) from APs. The difference between these arrival times (t4 - t1) provides a TDoA measurement. This avoids active participation, thus preserving privacy.
- Fingerprinting: To handle multipath and NLOS issues, PassiFi collects TDoA "fingerprints" at various reference points within the area of interest during an offline phase. These fingerprints are essentially TDoA measurements associated with specific locations.
- Deep Neural Network (DNN) for Localization: A deep learning model (specifically, a regression model with multiple fully connected hidden layers) is trained on these TDoA fingerprints. The DNN learns the complex, non-linear relationship between TDoA values and physical locations. This allows for more robust and accurate localization compared to traditional methods that might struggle with signal anomalies.
- Regularization Techniques: To prevent overfitting and improve generalization, the DNN employs dropout regularization (randomly dropping neurons during training) and early stopping (halting training when performance on a validation set plateaus).
- Online Smoothing Module: During the online localization phase, the system uses outlier rejection and temporal averaging to further refine the estimated location. This helps to smooth out noisy measurements and improve stability.
Datasets
The paper utilizes a large dataset collected at a real-world indoor testbed. This dataset is summarized in Table II and was collected using a device mounted on a robotic vehicle moving over approximately 5858 points. The testbed itself spans about 1500m² and features an office floor with metal frames and concrete walls, creating realistic multipath and NLOS conditions. The paper does not introduce a new benchmark dataset but rather uses an existing one for evaluation.
Key Insights and Ideas
- Privacy by Design: The most significant insight is that privacy can be inherently built into a localization system by shifting from active to passive signal measurement. This fundamentally changes the threat model, making it much harder for adversaries to track users.
- TDOA as a Robust Feature: TDoA measurements, when collected passively and fingerprinted, prove to be more resilient to common indoor signal issues (multipath, NLOS) than RSSI. This is because they are less susceptible to signal strength fluctuations.
- Deep Learning for Complex Mapping: The success of PassiFi highlights the power of deep neural networks in learning intricate spatial relationships from complex sensor data like TDoA fingerprints, surpassing traditional geometric or probabilistic approaches in challenging environments.
- Mitigating Active Protocol Vulnerabilities: By avoiding active FTM, PassiFi inherently bypasses known attacks like FTM frame spoofing and ACK frame spoofing, which exploit the active nature of RTT measurements.
Key Results
- Sub-meter Accuracy: PassiFi achieves a median location error of 0.82 meters.
- Significant Improvement over Multilateration: The system surpasses traditional multilateration by 128%.
- Competitive with State-of-the-Art: PassiFi's accuracy is on par with state-of-the-art active measurement systems.
- Privacy Preservation: The passive approach ensures users are not detected or tracked by adversaries.
- Robustness: The system demonstrates robustness to multipath and NLOS conditions through fingerprinting and deep learning.
Limitations and Future Work
- Initiating Station Placement: The paper notes that the location of the initiating station relative to responding stations can impact accuracy. Optimal placement, often near the center of mass of responding APs, is recommended.
- Scalability: The authors mention that assessing and boosting the system's scalability is a direction for future work.
- Dataset Limitations: While the dataset is large, it's from a specific testbed. Further evaluation across diverse environments would be beneficial.
- Tuning Parameters: The performance is sensitive to DNN hyperparameters (dropout rate, learning rate, number of layers), requiring careful tuning.
Summary updated