Private Fitness Tracking from Wearable Sensors with Zero-Knowledge Machine Learning

Abstract

Fitness apps help users track their exercises and health. Wearable sensors provide increasing amounts of data, bringing motivation but also challenges to privacy. Data shared by users’ devices may exceed what they want to share for the desired results. Health insurers, corporate wellness programs, and gamified apps all want proof of exercise, but current systems require users to disclose raw activity logs, heart rate, and location data. This paper presents a system that lets users prove they met exercise goals without revealing raw sensor data. We embed a Decision Tree classifier inside a zero-knowledge proof circuit using the Groth16 protocol. The classifier distinguishes rest, moderate, and vigorous activity from accelerometer and gyroscope readings. The model reaches 97.19% holdout accuracy on the MHEALTH dataset, validated through $\mathbf{1 0}$-fold crossvalidation. The circuit uses 10,447 constraints, small enough for smartphones and wearables. Proofs take about 500 ms on desktop and 1.1 seconds on mobile, with only 800 bytes each. Unlike systems where classification runs outside the proof, our approach makes the classification itself verifiable, so label injection attacks become computationally infeasible. We tested 30 cases covering all intensity classes and boundary conditions, and all proofs were generated and verified correctly. The system provides cryptographic guarantees while remaining practical for real-world deployment.