Researchers Improve Security For Smart Systems

In an increasingly connected and intelligent world, sensors collect and share vast amounts of data to help people make decisions.

Whether it's the smart grid constantly monitoring energy use or people trying to track health, people are receiving ever-increasing amounts of data in ways that are hard to decipher.

A group of WSU researchers recently developed a way to statistically analyze complex sensor data to make computer algorithms that make data-driven decisions more flexible and better able to handle small errors. The work has applications in many areas, including mobile health, smart homes, power grids and agriculture. They recently presented their work at the 2022 International Conference on Computer Aided Design in San Diego.

"This is an important and novel contribution to the field of machine learning system security," said Huey-Rogers Endowed Professor of Computer Science Jana Doppa, who led the work.

Machine learning algorithms are increasingly used for many applications such as smart grid management or smart agriculture. So, for example, they can collect data from farmland sensors and meteorological tools to learn and predict optimal irrigation times. For many smart applications, the collected data is represented as a time series, which is a collection of data that follows a sample over time and provides a series of timed data points.

Unfortunately, while computers gather information and produce these time series lines and graphs all day long, humans are not capable of reading and understanding them easily. More importantly, they can miss small but important changes, even those made maliciously.

"This kind of time series data is hard for people to understand," said Taha Belkhuja, a graduate student in the School of Electrical Engineering and Computer Science who has been working on the problem for three years. "If we introduce a small perturbation into this sequence of timestamps, you won't know whether we actually did the perturbation or not, and whether that reading is correct or not."

As machine learning is integrated into more and more systems, the security of those systems has been an understudied issue, Doppa said. Confusion attacks can occur when an attacker gains access to smart sensors and then causes small perturbations in the data in a way that is not apparent to the person viewing the data. Disruptions can cause forecasting and decision-making to fail.

"It's difficult from an interpretability standpoint, which means it's actually a huge advantage for an adversary who wants to hack some of that infrastructure," he said. "You cannot know what is right and what is wrong."

In their work, the WSU researchers introduced a security layer into their machine learning algorithm that looks for potential disturbances and determines the statistical probability of their occurrence, ensuring system resilience and preventing major failures. Working with wearable health monitoring devices, the researchers used their algorithm to automatically calculate irregularities in real sensor data and improve their accuracy by 50% compared to standard machine learning algorithms that they can't work with only clean data.

Because the algorithm analyzes the statistical probabilities of scenarios, the WSU team also saved computing power compared to traditional algorithms that must be recalculated over time. In particular, energy savings will result in less demand on the device's batteries.

In addition to Belkhuja and Doppa, researchers on this project included Ganapati Bhatt and Yan Yan, assistant professors in the School of Electrical Engineering and Computer Science, and Dina Hussain, a graduate student.

"Solving this complex problem requires expertise in multiple areas, and all members of the project team contributed synergistically," Doppa said.

Quote : Researchers improve security of smart systems (November 8, 2022) Retrieved November 11, 2022, from https://techxplore.com/news/2022-11-smart.html

This document is subject to copyright. Except in fair dealing for private study or research purposes, no part may be reproduced without written permission. The content is provided for informational purposes only.