Source-linked AI summary
Mobile Sensor Data Anonymization
Mohammad Malekzadeh, Richard G. Clegg, Andrea Cavallaro, Hamed Haddadi
TL;DR
Mobile and wearable motion data can expose user identity, creating a need for privacy-preserving release that retains application utility. The paper proposes an information-theoretic multi-objective autoencoder trained to suppress identity signals and limit distortion, including for unseen users. On activity-recognition data, the transformed data retained about 92% accuracy while user identification remained below 7%.
Problem
Motion data can reveal user-specific patterns and identity, while privacy-preserving release must limit this information without sacrificing service-specific utility.
Method
The paper trains an anonymizing autoencoder with an information-theoretic multi-objective loss that combines activity utility, distortion, and identity losses while shaping encoder and decoder outputs.
Results
Around 92% average activity-recognition accuracy was retained for raw and AAE data, while the AAE obscured user-identifiable information from a classifier trained on raw data.
Takeaways & Limitations
The trained anonymizer can be applied locally to new unseen users without user-specific retraining while preserving activity-recognition utility.
Takeaways & Limitations
Future work must measure the cost of local transformations, test other use cases, and derive statistical bounds for privacy protection.
Abstract
from arXiv · showhide
Motion sensors such as accelerometers and gyroscopes measure the instant acceleration and rotation of a device, in three dimensions. Raw data streams from motion sensors embedded in portable and wearable devices may reveal private information about users without their awareness. For example, motion data might disclose the weight or gender of a user, or enable their re-identification. To address this problem, we propose an on-device transformation of sensor data to be shared for specific applications, such as monitoring selected daily activities, without revealing information that enables user identification. We formulate the anonymization problem using an information-theoretic approach and propose a new multi-objective loss function for training deep autoencoders. This loss function helps minimizing user-identity information as well as data distortion to preserve the application-specific utility. The training process regulates the encoder to disregard user-identifiable patterns and tunes the decoder to shape the output independently of users in the training set. The trained autoencoder can be deployed on a mobile or wearable device to anonymize sensor data even for users who are not included in the training dataset. Data from 24 users transformed by the proposed anonymizing autoencoder lead to a promising trade-off between utility and privacy, with an accuracy for activity recognition above 92% and an accuracy for user identification below 7%.
1 INTRODUCTION
The paper targets on-device anonymization of mobile and wearable motion data, preserving application-specific utility while limiting user-identifying information. It proposes an information-theoretic, adversarially trained autoencoder designed to generalize beyond users in its training data.
- Motion sensor streams can reveal private information, including fine-grained behavioral profiles that expose user identity.
- The proposed mechanism transforms sensor data on the device so apps can receive service-relevant data without exposing unrelated sensitive information.
- The approach combines adversarial training with local anonymization, avoiding reliance on trusted aggregators or user participation in privacy-preserving training.
- The anonymization problem is formulated information-theoretically and addressed with a multi-objective loss for training deep autoencoders.
- The loss minimizes user-identity information while constraining distortion to preserve activity-recognition utility.
- The training process shapes both encoder features and decoder outputs so the resulting model can handle new users without user-specific retraining.
2 RELATED WORK
Prior work includes differential privacy, information-theoretic methods, and autoencoder-based transformations, but the paper emphasizes local deployment and protection of decoder outputs. Its approach is intended to generalize across unseen users without a trusted data aggregator.
- Adversarial learning can approximate data distributions, quantify mutual information, and remove sensitive information from latent representations.
- The authors assume access only to a public training dataset and require the anonymizer to generalize to unseen users.
- Autoencoders compress inputs into low-dimensional representations and reconstruct them, commonly by minimizing reconstruction differences.
- Corrupting or replacing inputs during training can encourage autoencoders to retain prominent patterns while ignoring noise or undesired patterns.
- Protecting only the encoder representation offers limited protection because the decoder may retain information extracted from training data.
- The paper therefore transforms outputs from both encoder and decoder and uses an adjustable distance constraint to control data distortion.
3 SENSOR DATA ANONYMIZATION
The paper formulates sensor anonymization as transforming motion data to suppress user identity information while preserving activity-recognition utility and limiting distortion. Its adversarially trained autoencoder regulates both latent representations and reconstructed outputs, producing transformed data that obscures user patterns while retaining activity differences.
- Anonymization function: The anonymization objective minimizes identity information in transformed data while maximizing activity information and minimizing distortion.The formulation uses mutual information for identity and activity together with a distance constraint between raw and transformed time series.
- Anonymization function: The AAE approximates this optimization by learning parameters for an autoencoder that maps input X to same-dimensional transformed output X′.The network is trained with adversarial objectives because searching over all possible anonymization functions is impractical.
- Architecture: The encoder suppresses user-identifiable information in low-dimensional Y, while decoder and activity regularizers evaluate reconstructed X′ for identity concealment and activity utility.EncReg receives Y and predicts user identity; DecReg receives X′ and predicts identity; ActReg receives X′ and predicts activity.
- Multi-objective loss function: The multi-objective loss combines identity, activity, and distortion terms, with non-negative weights controlling the utility-privacy trade-off.Identity loss helps remove user-specific signals, while activity and distortion losses can be customized to application requirements.
- Examples: The anonymized latent representation masks user clusters while preserving separation of Jogging from other activities.This behavior is shown for compressed gyroscope representations after transformation.
- Examples: The AAE obscures sensor patterns and peaks, introduces new periodic components, and maintains activity differences, reducing the possibility of re-identification from attributes such as height and weight.The comparisons cover gyroscope and accelerometer data across four activities and spectrograms of transformed data.
4 EVALUATION
The evaluation examines whether anonymization preserves activity-recognition utility while concealing user identity. Across sensor analyses and baselines, the AAE maintains activity performance while better obscuring identity-related patterns.
- Evaluation setup: The evaluation compares activity recognition and identity concealment, including Resampling, Singular Spectrum Analysis, and REP baselines.The study measures utility loss relative to raw data and evaluates transformed data under activity and identity recognition settings.
- Dataset: 24 participants performed six activities in 15 trials using accelerometer and gyroscope data collected by an iPhone 6s.The dataset uses controlled environments and two train-test splitting strategies.
- Sensor data characteristics: User weight affects autocorrelation peak intervals, creating identity information that downsampling cannot hide.For three users performing the same activity, the heaviest user exhibits the longest intervals between peaks.
- Baseline methods: Over 60% identity-recognition accuracy remains at 5Hz after downsampling, while SSA reconstructions using up to 10 components exceed 80% for both activity and identity recognition.These results show that baseline transformations can retain substantial identity information.
- AAE evaluation: AAE and raw data achieve around 92% average activity-recognition accuracy under the Subject setting.The transformed data preserve utility and can slightly improve average activity-recognition performance compared with raw data.
- AAE evaluation: The AAE obscures identity information more effectively than downsampling, even against a classifier trained on users’ raw data.The classifier cannot distinguish users reliably after the AAE transformation, while downsampling from 50Hz to 5Hz reveals more information.
5 CONCLUSION
The paper concludes that its anonymizing autoencoder balances privacy and utility for mobile and wearable sensor data. It also identifies deployment cost, broader use cases, and formal privacy guarantees as future work.
- Conclusion: The proposed multi-objective loss trains an anonymizing autoencoder that removes user-identifiable features while minimally perturbing sensor data.The decoder is shaped independently of training users, supporting use with unseen users and service-specific activity recognition.
- Conclusion: The approach is intended to support anonymization in participatory sensing for health and well-being data analysis.The paper frames this as a use case for individuals contributing data from personal devices.
- Future work: Future work will measure on-device transformation cost, test other tasks, and derive statistical bounds for achieved privacy protection.These are stated as directions for extending the evaluation and formalizing privacy guarantees.