Hardware Security Enhancement in Cyber-Physical Systems using Deep Learning-based Anomaly Detection
| Title | Hardware Security Enhancement in Cyber-Physical Systems using Deep Learning-based Anomaly Detection |
|---|---|
| Summary | In this project, we intend to employ a deep learning approach to detect anomalies in cyber-physical systems using data flow monitoring. |
| Keywords | Cyber-Physical Systems, Anomaly Detection, Hardware Security, Deep Learning, Low Power |
| TimeFrame | |
| References | |
| Prerequisites | |
| Author | Mahdi Fazeli |
| Supervisor | Mahdi Fazeli |
| Level | Master |
| Status | Open |
The use of third-party IP cores and or Commercial off the Shelf (COTS) components have become a common practice in the design of embedded systems as it significantly reduces the development cost and time to market. However, the possible presence of hardware Trojans in such components may result in serious security challenges specially. Functional testing is not an efficient solution as the hardware Trojan circuits are rarely activated. Nowadays embedded Cyber-Physical systems (CPS), on the other hand, are playing a serious role in our daily life. CPS as a class of embedded systems is designed to perform specific and few tasks. These systems cautiously interact with the physical environments through sensors and actuators. A CPS briefly consists of four main parts including the computation part or the Cyber part, the physical environment, sensors, and actuators. The input and output data are the most important entities in a CPS. Data in a CPS can be classified as numeric and categorical data. Numeric data are continuous values with some mathematical ordering properties, e.g. the difference and ratio of such values are meaningful. Categorical data are series of unit labels or semantic information with no mathematical ordering e.g. temperature on a summer day cannot be -10. Using mathematical ordering and semantic information can be efficiently used to detect anomalous behavior of CPS due to the presence of possible hardware Trojans. This approach is feasible in embedded systems as designers have a high level of information regarding the behavior of such systems. As such information is extracted, the probability of capturing an anomaly would significantly increase. In this project, we intend to employ a deep learning approach to detect anomalies in CPS based on the mentioned features. In the first step,
for a specific CPS e.g a medical CPS, we should carefully extract as much numerical and categorical data as possible. In the second step, we should employ a cost and energy-efficient deep learning approach and a neural network to be trained by the mentioned data. In the final step, we should evaluate the approach in the execution time and retrain the neural network in case of having an accuracy less than an expected threshold. The main and important parts of this project are 1) careful extraction of numerical and categorical data and corresponding ordering and semantics to have maximum accuracy in detecting the anomalous behavior of the under-test CPS, and 2) designing lightweight hardware in terms of power consumption and cost for the anomaly detection system.
