Search for: All records

Audio-based human activity recognition (HAR) is very popular because many human activities have unique sound signatures that can be detected using machine learning (ML) approaches. These audio-based ML HAR pipelines often use common featurization techniques, such as extracting various statistical and spectral features by converting time domain signals to the frequency domain (using an FFT) and using them to train ML models. Some of these approaches also claim privacy benefits by preventing the identification of human speech. However, recent deep learning-based automatic speech recognition (ASR) models pose new privacy challenges to these featurization techniques. In this paper, we systematically evaluate various featurization approaches for audio data, assessing their privacy risks through metrics like speech intelligibility (PER and WER) while considering the utility tradeoff in terms of ML-based activity recognition accuracy. Our findings reveal the susceptibility of these approaches to speech content recovery when exposed to recent ASR models, especially under re-tuning or retraining conditions. Notably, fine-tuned ASR models achieved an average Phoneme Error Rate (PER) of 39.99% and Word Error Rate (WER) of 44.43% in speech recognition for these approaches. To overcome these privacy concerns, we propose Kirigami, a lightweight machine learning-based audio speech filter that removes human speech segments reducing the efficacy of ASR models (70.48% PER and 101.40% WER) while also maintaining HAR accuracy (76.0% accuracy). We show that Kirigami can be implemented on common edge microcontrollers with limited computational capabilities and memory, providing a path to deployment on a variety of IoT devices. Finally, we conducted a real-world user study and showed the robustness of Kirigami on a laptop and an ARM Cortex-M4F microcontroller under three different background noises. 

Translating fine-grained activity detection (e.g., phone ring, talking interspersed with silence and walking) into semantically meaningful and richer contextual information (e.g., on a phone call for 20 minutes while exercising) is essential towards enabling a range of healthcare and human-computer interaction applications. Prior work has proposed building ontologies or temporal analysis of activity patterns with limited success in capturing complex real-world context patterns. We present TAO, a hybrid system that leverages OWL-based ontologies and temporal clustering approaches to detect high-level contexts from human activities. TAO can characterize sequential activities that happen one after the other and activities that are interleaved or occur in parallel to detect a richer set of contexts more accurately than prior work. We evaluate TAO on real-world activity datasets (Casas and Extrasensory) and show that our system achieves, on average, 87% and 80% accuracy for context detection, respectively. We deploy and evaluate TAO in a real-world setting with eight participants using our system for three hours each, demonstrating TAO's ability to capture semantically meaningful contexts in the real world. Finally, to showcase the usefulness of contexts, we prototype wellness applications that assess productivity and stress and show that the wellness metrics calculated using contexts provided by TAO are much closer to the ground truth (on average within 1.1%), as compared to the baseline approach (on average within 30%). 

Modern Internet of Things (IoT) applications, from contextual sensing to voice assistants, rely on ML-based training and serving systems using pre-trained models to render predictions. However, real-world IoT environments are diverse, with rich IoT sensors and need ML models to be personalized for each setting using relatively less training data. Most existing general-purpose ML systems are optimized for specific and dedicated hardware resources and do not adapt to changing resources and different IoT application requirements. To address this gap, we propose MLIoT, an end-to-end Machine Learning System tailored towards supporting the entire lifecycle of IoT applications. MLIoT adapts to different IoT data sources, IoT tasks, and compute resources by automatically training, optimizing, and serving models based on expressive applicationspecific policies. MLIoT also adapts to changes in IoT environments or compute resources by enabling re-training, and updating models served on the fly while maintaining accuracy and performance. Our evaluation across a set of benchmarks show that MLIoT can handle multiple IoT tasks, each with individual requirements, in a scalable manner while maintaining high accuracy and performance. We compare MLIoT with two state-of-the-art hand-tuned systems and a commercial ML system showing that MLIoT improves accuracy from 50% - 75% while reducing or maintaining latency. 

There is increasing interest in deploying building-scale, general-purpose, and high-fidelity sensing to drive emerging smart building applications. However, the real-world deployment of such systems is challenging due to the lack of system and architectural support. Most existing sensing systems are purpose-built, consisting of hardware that senses a limited set of environmental facets, typically at low fidelity and for short-term deployment. Furthermore, prior systems with high-fidelity sensing and machine learning fail to scale effectively and have fewer primitives, if any, for privacy and security. For these reasons, IoT deployments in buildings are generally short-lived or done as a proof of concept. We present the design of Mites, a scalable end-to-end hardware-software system for supporting and managing distributed general-purpose sensors in buildings. Our design includes robust primitives for privacy and security, essential features for scalable data management, as well as machine learning to support diverse applications in buildings. We deployed our Mites system and 314 Mites devices in Tata Consultancy Services (TCS) Hall at Carnegie Mellon University (CMU), a fully occupied, five-story university building. We present a set of comprehensive evaluations of our system using a series of microbenchmarks and end-to-end evaluations to show how we achieved our stated design goals. We include five proof-of-concept applications to demonstrate the extensibility of the Mites system to support compelling IoT applications. Finally, we discuss the real-world challenges we faced and the lessons we learned over the five-year journey of our stack's iterative design, development, and deployment.