Evaluating Attribute Inference Risks in Urban Care Taxi Arrival Time Prediction Models Using Geospatial Data

Research Note

Care services for older adults are becoming part of urban infrastructure, and machine learning models trained on location data are being considered for optimizing how care taxis are dispatched, in programmes such as Adult Day Care Transportation in the US and Total Mobility in New Zealand. Because such data describes people's daily lives in detail, sharing it between cities is difficult, which is why frameworks like federated learning have been proposed to allow collaboration without moving the data itself.

Withholding the data does not by itself remove the risk. An attribute inference attack queries a shared model in order to recover attributes of the individuals whose records were used to train it. A model absorbs the tendencies of its training data, so careful observation of how it responds can reveal information that was never meant to leave its owner. What makes this awkward is that the data was never handed over, yet its contents can still be estimated.

This study builds a boarding and alighting time predictor from real-world care taxi traces and then applies an attribute inference attack to measure how accurately an attacker can infer whether a given user has a walking disability.

It also examines how data processing intended to improve the model affects vulnerability to that attack, and the result is instructive. When class imbalance was addressed using SMOTE data augmentation, the accuracy of the attack rose from 61.3% to 73.0%. SMOTE synthesises additional minority-class samples by interpolation and is used routinely to improve predictive performance, but it also impresses the characteristics of the minority class more firmly onto the model, which works in the direction of making the people in that minority easier to single out.

Preprocessing chosen to raise accuracy can therefore become a privacy weakness in its own right. The findings offer guidance on what to watch for when designing machine learning systems in a domain as sensitive as eldercare transportation.