Search for: All records

Machine learning algorithms are increasingly used for inference and decision-making in embedded systems. Data from sensors are used to train machine learning models for various smart functions of embedded and cyber-physical systems ranging from applications in healthcare, autonomous vehicles, and national security. However, recent studies have shown that machine learning models can be fooled by adding adversarial noise to their inputs. The perturbed inputs are called adversarial examples. Furthermore, adversarial examples designed to fool one machine learning system are also often effective against another system. This property of adversarial examples is calledadversarial transferabilityand has not been explored in wearable systems to date. In this work, we take the first stride in studying adversarial transferability in wearable sensor systems from four viewpoints: (1) transferability between machine learning models; (2) transferability across users/subjects of the embedded system; (3) transferability across sensor body locations; and (4) transferability across datasets used for model training. We present a set of carefully designed experiments to investigate these transferability scenarios. We also propose a threat model describing the interactions of an adversary with the source and target sensor systems in different transferability settings. In most cases, we found high untargeted transferability, whereas targeted transferability success scores varied from 0% to 80%. The transferability of adversarial examples depends on many factors such as the inclusion of data from all subjects, sensor body position, number of samples in the dataset, type of learning algorithm, and the distribution of source and target system dataset. The transferability of adversarial examples decreased sharply when the data distribution of the source and target system became more distinct. We also provide guidelines and suggestions for the community for designing robust sensor systems. 

New modes of technology are offering unprecedented opportunities to unobtrusively collect data about people's behavior. While there are many use cases for such information, we explore its utility for predicting multiple clinical assessment scores. Because clinical assessments are typically used as screening tools for impairment and disease, such as mild cognitive impairment (MCI), automatically mapping behavioral data to assessment scores can help detect changes in health and behavior across time. In this article, we aim to extract behavior markers from two modalities, a smart home environment and a custom digital memory notebook app, for mapping to 10 clinical assessments that are relevant for monitoring MCI onset and changes in cognitive health. Smart-home-based behavior markers reflect hourly, daily, and weekly activity patterns, while app-based behavior markers reflect app usage and writing content/style derived from free-form journal entries. We describe machine learning techniques for fusing these multimodal behavior markers and utilizing joint prediction. We evaluate our approach using three regression algorithms and data from 14 participants with MCI living in a smart-home environment. We observed moderate to large correlations between predicted and ground-truth assessment scores, ranging from r = 0.601 to r = 0.871 for each clinical assessment. 

Abstract Background Behavior and health are inextricably linked. As a result, continuous wearable sensor data offer the potential to predict clinical measures. However, interruptions in the data collection occur, which create a need for strategic data imputation. Objective The objective of this work is to adapt a data generation algorithm to impute multivariate time series data. This will allow us to create digital behavior markers that can predict clinical health measures. Methods We created a bidirectional time series generative adversarial network to impute missing sensor readings. Values are imputed based on relationships between multiple fields and multiple points in time, for single time points or larger time gaps. From the complete data, digital behavior markers are extracted and are mapped to predicted clinical measures. Results We validate our approach using continuous smartwatch data for n = 14 participants. When reconstructing omitted data, we observe an average normalized mean absolute error of 0.0197. We then create machine learning models to predict clinical measures from the reconstructed, complete data with correlations ranging from r = 0.1230 to r = 0.7623. This work indicates that wearable sensor data collected in the wild can be used to offer insights on a person's health in natural settings.