Deep learning (DL) has demonstrated great performance in various applications on powerful computers and servers. Recently, with the advancement of more powerful mobile devices (e.g., smartphones and touch pads), researchers are seeking DL solutions that could be deployed on mobile devices. Compared to traditional DL solutions using cloud servers, deploying DL on mobile devices have unique advantages in data privacy, communication overhead, and system cost. This article provides a comprehensive survey for the current studies of adopting and deploying DL on mobile devices. Specifically, we summarize and compare the state-of-the-art DL techniques on mobile devices in various application domains involving vision, speech/speaker recognition, human activity recognition, transportation mode detection, and security. We generalize an optimization pipeline for bringing DL to mobile devices, including model-oriented optimization mechanisms (e.g., pruning and quantization) and nonmodel-oriented optimization mechanisms (e.g., software accelerator and hardware design). Moreover, we summarize popular DL libraries regarding their support to state-of-the-art models (software) and processors (hardware). Based on our summarization, we further provide insights into potential research opportunities for developing DL for mobile devices.

Traditional one-time user authentication processes might cause friction and unfavorable user experience in many widely-used applications. This is a severe problem in particular for security-sensitive facilities if an adversary could obtain unauthorized privileges after a user’s initial login. Recently, continuous user authentication (CA) has shown its great potential by enabling seamless user authentication with few active participation. We devise a low-cost system exploiting a user’s pulsatile signals from the photoplethysmography (PPG) sensor in commercial wrist-worn wearables for CA. Compared to existing approaches, our system requires zero user effort and is applicable to practical scenarios with non-clinical PPG measurements having motion artifacts (MA). We explore the uniqueness of the human cardiac system and design an MA filtering method to mitigate the impacts of daily activities. Furthermore, we identify general fiducial features and develop an adaptive classifier using the gradient boosting tree (GBT) method. As a result, our system can authenticate users continuously based on their cardiac characteristics so little training effort is required. Experiments with our wrist-worn PPG sensing platform on 20 participants under practical scenarios demonstrate that our system can achieve a high CA accuracy of over 90% and a low false detection rate of 4% in detecting random attacks.

This paper subverts the traditional understanding of Photoplethysmography (PPG) and opens up a new direction of the utility of PPG in commodity wearable devices, especially in the domain of human computer interaction of fine-grained gesture recognition. We demonstrate that it is possible to leverage the widely deployed PPG sensors in wrist-worn wearable devices to enable finger-level gesture recognition, which could facilitate many emerging human-computer interactions (e.g., sign-language interpretation and virtual reality). While prior solutions in gesture recognition require dedicated devices (e.g., video cameras or IR sensors) or leverage various signals in the environments (e.g., sound, RF or ambient light), this paper introduces the first PPG-based gesture recognition system that can differentiate fine-grained hand gestures at finger level using commodity wearables. Our innovative system harnesses the unique blood flow changes in a user's wrist area to distinguish the user's finger and hand movements. The insight is that hand gestures involve a series of muscle and tendon movements that compress the arterial geometry with different degrees, resulting in significant motion artifacts to the blood flow with different intensity and time duration. By leveraging the unique characteristics of the motion artifacts to PPG, our system can accurately extract the gesture-related signals from the significant background noise (i.e., pulses), and identify different minute finger-level gestures. Extensive experiments are conducted with over 3600 gestures collected from 10 adults. Our prototype study using two commodity PPG sensors can differentiate nine finger-level gestures from American Sign Language with an average recognition accuracy over 88%, suggesting that our PPG-based finger-level gesture recognition system is promising to be one of the most critical components in sign language translation using wearables.

Different from traditional gestures, sign language gestures involve a lot of finger-level gestures without wrist or arm movements. They are hard to detect using existing motion sensors-based approaches. We introduce the first low-cost sign language gesture recognition system that can differentiate fine-grained finger movements using the Photoplethysmography (PPG) and motion sensors in commodity wearables. By leveraging the motion artifacts in PPG, our system can accurately recognize sign language gestures when there are large body movements, which cannot be handled by the traditional motion sensor-based approaches. We further explore the feasibility of using both PPG and motion sensors in wearables to improve the sign language gesture recognition accuracy when there are limited body movements. We develop a gradient boost tree (GBT) model and deep neural network-based model (i.e., ResNet) for classification. The transfer learning technique is applied to ResNet-based model to reduce the training effort. We develop a prototype using low-cost PPG and motions sensors and conduct extensive experiments and collect over 7000 gestures from 10 adults in the static and body-motion scenarios. Results demonstrate that our system can differentiate nine finger-level gestures from the American Sign Language with an average recognition accuracy over 98 percent.

The ever-growing security issues in various mobile applications and smart devices create an urgent demand for a reliable and convenient user verification method. Traditional verification methods request users to provide their secrets (e.g., entering passwords and collecting fingerprints). We envision that the essential trend of user verification is to free users from active participation in the verification process. Toward this end, we propose a continuous user verification system, which re-uses the widely deployed WiFi infrastructure to capture the unique physiological characteristics rooted in user’s respiratory motions. Different from the existing continuous verification approaches, posing dependency on restricted scenarios/user behaviors (e.g., keystrokes and gaits), our system can be easily integrated into any WiFi infrastructure to provide non-intrusive continuous verification. Specifically, we extract the respiration-related signals from the channel state information (CSI) of WiFi. We then derive the user-specific respiratory features based on the waveform morphology analysis and fuzzy wavelet transformation of the respiration signals. Additionally, a deep learning based user verification scheme is developed to identify legitimate users accurately and detect the existence of spoofing attacks. Extensive experiments involving 20 participants demonstrate that the proposed system can robustly verify/identify users and detect spoofers under various types of attacks.

Deep learning models have become key enablers of voice user interfaces. With the growing trend of adopting outsourced training of these models, backdoor attacks, stealthy yet effective training-phase attacks, have gained increasing attention. They inject hidden trigger patterns through training set poisoning and overwrite the model's predictions in the inference phase. Research in backdoor attacks has been focusing on image classification tasks, while there have been few studies in the audio domain. In this work, we explore the severity of audio-domain backdoor attacks and demonstrate their feasibility under practical scenarios of voice user interfaces, where an adversary injects (plays) an unnoticeable audio trigger into live speech to launch the attack. To realize such attacks, we consider jointly optimizing the audio trigger and the target model in the training phase, deriving a position-independent, unnoticeable, and robust audio trigger. We design new data poisoning techniques and penalty-based algorithms that inject the trigger into randomly generated temporal positions in the audio input during training, rendering the trigger resilient to any temporal position variations. We further design an environmental sound mimicking technique to make the trigger resemble unnoticeable situational sounds and simulate played over-the-air distortions to improve the trigger's robustness during the joint optimization process. Extensive experiments on two important applications (i.e., speech command recognition and speaker recognition) demonstrate that our attack can achieve an average success rate of over 99% under both digital and physical attack settings.