Accurate, Energy-efficient, and Secure Machine Learning Models: Applications to Smart Healthcare

Abstract

Machine learning (ML) algorithms automate the data-label mapping and decision-making processes. The output of the ML algorithms determines the next step in real-world applications. If the ML output is diagnosis of a disease, e.g., in smart healthcare applications, then the next step is the treatment options. On the other hand, if the ML output is the traffic condition, e.g., in smart city applications, then the next step is traffic regulation. It is desirable to have ML-based systems that are simultaneously accurate, secure, energy-efficient, low-cost, silently operable, maintainable, customizable, low-delay, and scalable. However, there are tradeoffs among these design goals that impact the complexity of the overall system. This thesis focuses on the accuracy, security, and energy-efficiency objectives.

We first target the accuracy objective with a new dual-space (feature and semantic space) classification approach: SECRET. While traditional supervised learning approaches operate in the feature space only, SECRET utilizes both feature and semantic spaces in the classification process. It incorporates class affinity and dissimilarity information into the decision process using the semantic space. This property enables SECRET to make informed decisions on class labels, thus enhancing its overall classification performance. By demonstrating the significant benefits of the semantic space, SECRET opens up a new research direction for the machine learning community.

Next, we introduce an automatic stress detection and alleviation system: SoDA. SoDA takes advantage of emerging wearable medical sensors to continuously monitor human stress levels and mitigate stress as it arises. It performs stress detection and alleviation in a user-transparent manner. When it detects stress, SoDA employs a stress alleviation technique in an adaptive manner based on the stress response of the user.

Furthermore, we present machine learning models that are built with smartphone and wearable medical sensor data and enable the smartphone to understand us. We call this system your smartphone understands you (YSUY). By understanding our physical, mental, and emotional states, YSUY promotes quality of life of the users by assessing their states and shining light on fundamental human-centric needs.

Finally, we introduce a simultaneously smart, secure, and energy-efficient Internet-of-Things (IoT) sensor architecture: SSE. In SSE, we use inference in the compressively-sensed domain and transmit data to the base station when an event of interest occurs. Since on-sensor compression and inference drastically reduce the amount of data that need to be transmitted, we are able to reduce the IoT sensor energy by two-to-three orders of magnitude. A small part of this energy bonus is used to carry out encryption and hashing to ensure data confidentiality and integrity. Overall, we achieve smartness through decision-making inferences, security through encryption and hashing, and energy efficiency through both compression and decision-making inferences. By performing data compression and machine learning inference on the IoT sensor node, the SSE approach not only enables the IoT system to push signal processing and decision-making to the extreme of the edge-side (i.e., the sensor node), but also solves the data security and energy efficiency problems simultaneously.