University Events

Description: Topic: Decentralized Reinforcement Learning with FPGA Acceleration for Anomaly Detection in UAV Network

Location: Science & Engineering Building (SENG), Room 212

Zoom Conference Link: https://umassd.zoom.us/j/98135954802
Meeting ID: 981 3595 4802
Passcode: 759903

Abstract:
The field of cybersecurity is continually growing as new and more adaptive threats rise. Not only this, but as the Internet of Things (IOT) becomes more widespread, devices that need security also further diversifies. This thesis explores the use of machine learning as a method to adapt to new cybersecurity threats for unmanned aerial vehicle (UAV) security, without the need for centralization. In the proposed model, a UAV serves as a relay between ground users and a cloud server. Ground users send data through the UAV relay to the cloud server, with the possibility of malicious users also sending data. The UAV provides an intrusion detection system (IDS) for the cloud server, blocking or forwarding packets based on a reinforcement learning Deep-Q Network (DQN). The network used was trained on the CICIDS2017 data set, with a classification accuracy of 86.5% across six classes, and an anomaly detection accuracy of 99.9%. Additionally, the UAV was equipped with a convolutional neural network (CNN) to detect jamming attacks that may occur on the UAV's frequency band. This model was trained on a global navigation satellite system (GNSS) jamming data set with six classes - achieving a classification accuracy of 88.1% and anomaly detection rate of 99.1%. By achieving such high detection rates, this negates the need for a centralized UAV model as all UAVs can be self-sustained. Both the CNN and DQN were then compiled to run on an Zynq UltraSCALE+ ZCU102 system-on-chip (SoC) field-programmable gate array (FPGA) using Xilinx's Vitis-AI software. This simulates the possibility of using a Xilinx's Vitis-AI hardware to speed up machine learning inference time. The compiled model provides a speed up in inference of 9.67x and 47.1x on the CNN and DQN respectively when compared to an AMD Ryzen 5600x CPU, with a minimal impact on accuracy. These results demonstrate that deep learning can effectively be used to provide UAV security as both an IDS for a cloud server and jamming detection system for itself. It also showcases the performance benefits of using FPGA hardware acceleration to maximize the throughput of the UAV.

Advisor(s): Dr. Ruolin Zhou, Assistant Professor, Department of Electrical & Computer Engineering, UMASS Dartmouth

Committee Members: Dr. Hong Liu, Professor, Department of Electrical & Computer Engineering, UMASS Dartmouth; Dr. Honggang Wang, Professor, Department of Electrical & Computer Engineering, UMASS Dartmouth

Note: All ECE Graduate Students are ENCOURAGED to attend.
All interested parties are invited to attend. Open to the public.

*For further information, please contact Dr. Ruolin Zhou via email at rzhou1@umassd.edu