Billions of distributed, heterogeneous, and resource constrained IoT devices deploy on-device machine learning (ML) for private, fast, and offline inference on personal data. On-device ML is highly context dependent and sensitive to user, usage, hardware, and environment attributes. This sensitivity and the propensity toward bias in ML makes it important to study bias in on-device settings. Our study is one of the first investigations of bias in this emerging domain and lays important foundations for building fairer on-device ML. We apply a software engineering lens, investigating the propagation of bias through design choices in on-device ML workflows. We first identify reliability bias as a source of unfairness and propose a measure to quantify it. We then conduct empirical experiments for a keyword spotting task to show how complex and interacting technical design choices amplify and propagate reliability bias. Our results validate that design choices made during model training, like the sample rate and input feature type, and choices made to optimize models, like light-weight architectures, the pruning learning rate, and pruning sparsity, can result in disparate predictive performance across male and female groups. Based on our findings, we suggest low effort strategies for engineers to mitigate bias in on-device ML.

@en

When deploying machine learning (ML) models on embedded and IoT devices, performance encompasses more than an accuracy metric: inference latency, energy consumption, and model fairness are necessary to ensure reliable performance under heterogeneous and resource-constrained operating conditions. To this end, prior research has studied model-centric approaches, such as tuning the hyperparameters of the model during training and later applying model compression techniques to tailor the model to the resource needs of an embedded device. In this paper, we take a data-centric view of embedded ML and study the role that pre-processing parameters in the data pipeline can play in balancing the various performance metrics of an embedded ML system. Through an in-depth case study with audio-based keyword spotting (KWS) models, we show that pre-processing parameter tuning is a remarkable tool that model developers can adopt to trade-off between a model's accuracy, fairness, and system efficiency, as well as to make an embedded ML model resilient to unseen deployment conditions.@en

Wearable sensors are increasingly becoming the primary interface for monitoring human activities. However, in order to scale human activity recognition (HAR) using wearable sensors to million of users and devices, it is imperative that HAR computational models are robust against real-world heterogeneity in inertial sensor data. In this paper, we study the problem of wearing diversity which pertains to the placement of the wearable sensor on the human body, and demonstrate that even state-of-the-art deep learning models are not robust against these factors. The core contribution of the paper lies in presenting a first-of-its-kind in-depth study of unsupervised domain adaptation (UDA) algorithms in the context of wearing diversity - we develop and evaluate three adaptation techniques on four HAR datasets to evaluate their relative performance towards addressing the issue of wearing diversity. More importantly, we also do a careful analysis to learn the downsides of each UDA algorithm and uncover several implicit data-related assumptions without which these algorithms suffer a major degradation in accuracy. Taken together, our experimental findings caution against using UDA as a silver bullet for adapting HAR models to new domains, and serve as practical guidelines for HAR practitioners as well as pave the way for future research on domain adaptation in HAR.

@en

The increasing availability of multiple sensory devices on or near a human body has opened brand new opportunities to leverage redundant sensory signals for powerful sensing applications. For instance, personal-scale sensory inferences with motion and audio signals can be done individually on a smartphone, a smartwatch, and even an earbud - each offering unique sensor quality, model accuracy, and runtime behaviour. At execution time, however, it is incredibly challenging to assess these characteristics to select the best device for accurate and resource-efficient inferences. To this end, we look at a quality-aware collaborative sensing system that actively interplays across multiple devices and respective sensing models. It dynamically selects the best device as a function of model accuracy at any given context. We propose two complementary techniques for the runtime quality assessment. Borrowing principles from active learning, our first technique runs on three heuristic-based quality assessment functions that employ confidence, margin sampling, and entropy of models' output. Our second technique is built with a siamese neural network and acts on the premise that runtime sensing quality can be learned from historical data. Our evaluation across multiple motion and audio datasets shows that our techniques provide 12% increase in overall accuracy through dynamic device selection at the average expense of 13 mW power on each device as compared to traditional single-device approaches.@en

Conversational agents are increasingly becoming digital partners of our everyday computing experiences offering a variety of purposeful information and utility services. Although rich on competency, these agents are entirely oblivious to their users' situational and emotional context today and incapable of adjusting their interaction style and tone contextually. To this end, we present a mixed-method study that informs the design of a situation-and emotion-aware conversational agent for kinetic earables. We surveyed 280 users, and qualitatively interviewed 12 users to understand their expectation from a conversational agent in adapting the interaction style. Grounded on our findings, we develop a first-of-its-kind emotion regulator for a conversational agent on kinetic earable that dynamically adjusts its conversation style, tone, volume in response to users emotional, environmental, social and activity context gathered through speech prosody, motion signals and ambient sound. We describe these context models, the end-to-end system including a purpose-built kinetic earable and their real-world assessment. The experimental results demonstrate that our regulation mechanism invariably elicits better and affective user experience in comparison to baseline conditions in different real-world settings.

@en

We explore a new variability observed in motion signals acquired from modern wearables. Wearing variability refers to the variations of the device orientation and placement across wearing events. We collect the accelerometer data on a smartwatch and an earbud and analyse how motion signals change due to the wearing variability. Our analysis shows that the wearing variability can bring an unexpected change to motion signals, not only from different users but also from different wearing sessions of the same user. We also provide empirical ranges of changes in device orientations.

@en

In this paper, we introduce inertial signals obtained from an earable placed in the ear canal as a new compelling sensing modality for recognising two key facial expressions: Smile and frown. Borrowing principles from Facial Action Coding Systems, we first demonstrate that an inertial measurement unit of an earable can capture facial muscle deformation activated by a set of temporal microexpressions. Building on these observations, we then present three different learning schemes - shallow models with statistical features, hidden Markov model, and deep neural networks to automatically recognise smile and frown expressions from inertial signals. The experimental results show that in controlled non-conversational settings, we can identify smile and frown with high accuracy (F1 score: 0.85).

@en

Deep learning has enabled personal and IoT devices to rethink microphones as a multi-purpose sensor for understanding conversation and the surrounding environment. This resulted in a proliferation of Voice Controllable Systems (VCS) around us. The increasing popularity of such systems is also prone to attracting miscreants, who often want to take advantage of the VCS without the knowledge of the user. Consequently, understanding the robustness of VCS, especially under adversarial attacks, has become an important research topic. Although there exists some previous work on audio adversarial attacks, their scopes are limited to embedding the attacks onto pre-recorded music clips, which when played through speakers cause VCS to misbehave. As an attack-audio needs to be played, the occurrence of this type of attacks can be suspected by a human listener. In this paper, we focus on audio-based Denial-of-Service (DoS) attack, which is unexplored in the literature. Contrary to previous work, we show that adversarial audio attacks in real-time and overthe-air are possible, while a user interacts with VCS. We show that the attacks are effective regardless of the user's command and interaction timings. In this paper, we present a first-of-itskind imperceptible and always-on universal audio perturbation technique that enables such DoS attack to be successful. We thoroughly evaluate the performance of the attacking scheme across (i) two learning tasks, (ii) two model architectures and (iii) three datasets. We demonstrate that the attack can introduce as high as 78% error rate in audio recognition tasks.

@en

Despite signicant advances in the performance of sensory inference models, their poor robustness to changing environmental conditions and hardware remains a major hurdle for widespread adoption. In this paper, we introduce the concept of unsupervised domain adaptation which is a technique to adapt sensory inference models to new domains only using unlabeled data from the target domain. We present two case-studies to motivate the problem and highlight some of our recent work in this space. Finally, we discuss the core challenges in this space that can trigger further ubicomp research on this topic.@en

We present eSense - an open and multi-sensory in-ear wearable platform for personal-scale behaviour analytics. eSense is a true wireless stereo (TWS) earbud and supports dual-mode Bluetooth and Bluetooth Low Energy. It is also augmented with a 6-axis inertial measurement unit and a microphone. We demonstrate the eSense platform, the data exploration tool with the open APIs for the real-time visualisation of multi-modal sensory data, and its manifestation in a 360 ^◦ workplace well-being application.

@en

We present eSense - an open and multi-sensory in-ear wearable platform to detect and monitor human activities. eSense is a true wireless stereo (TWS) earbud with dual-mode Bluetooth and Bluetooth Low Energy and augmented with a 6-axis inertial measurement unit and a microphone. We showcase the eSense platform, its data APIs to capture real-time multi-modal sensory data in a data exploration tool, and its manifestation in a 360◦ workplace well-being application.@en

We propose a cross-modal approach for conversational well-being monitoring with a multi-sensory earable. It consists of motion, audio, and BLE models on earables. Using the IMU sensor, the microphone, and BLE scanning, the models detect speaking activities, stress and emotion, and participants in the conversation, respectively. We discuss the feasibility in qualifying conversations with our purpose-built cross-modal model in an energy-efficient and privacy-preserving way. With the cross-modal model, we develop a mobile application that qualifies on-going conversations and provides personalised feedback on social well-being.

@en

In this paper, we explore audio and kinetic sensing on earable devices with the commercial on-the-shelf form factor. For the study, we prototyped earbud devices with a 6-axis inertial measurement unit and a microphone. We systematically investigate the differential characteristics of the audio and inertial signals to assess their feasibility in human activity recognition. Our results demonstrate that earable devices have a superior signal-to-noise ratio under the influence of motion artefacts and are less susceptible to acoustic environment noise. We then present a set of activity primitives and corresponding signal processing pipelines to showcase the capabilities of earbud devices in converting accelerometer, gyroscope, and audio signals into the targeted human activities with a mean accuracy reaching up to 88% in varying environmental conditions.

@en