Next-generation (NextG) cellular systems will be designed with awareness, intelligence, and flexibility to support diverse use cases such as telepresence, immersive sports, connected intelligent machines and interacting robots, precision healthcare, and others. These attributes will be realized through groundbreaking technologies in uncharted frequency bands, e.g., millimeter-wave (mmWave) and Teraherz (THz) bands, virtualized core and radio access network (RAN) architectures, new spectrum sharing models, and powerful machine learning algorithms that optimally manage resources. Various AI/ML approaches can be adopted for NextG RAN management and control, such as reinforcement learning (RL). However, RAN resource sharing exposes the network to new security threats that target the robustness of the decision-making process. Our research aims to study RAN resource management algorithms in an adversarial setting, mitigate user privacy leakage, and develop mechanisms to ensure that RAN policies are designed to meet service level agreements.

To meet the diverse service demands of NextG applications, NextG networks will support new wireless capabilities in the mmWave and THz bands that go beyond communications and simultaneously support high-resolution sensing. By integrating sensing into the communications network, the network acts as a "radar" sensor, using its own radio signals to sense and comprehend the physical world in which it operates. The sensing data can then be leveraged to enhance the network’s own operations, augment existing services such as XR and digital twinning, and enable new services such as gesture/activity recognition, imaging and environment reconstruction. While ISAC offers significant performance benefits, its security and resiliency issues have been largely under-explored. Our research investigates the security of ISAC including novel vulnerabilities and defense mechanisms, and exploit ISAC to build secure-by-design NextG applications.

Autonomous Systems (AS), such as self-driving cars, robotic agents (Embodied AI), and surveillance systems, rely heavily on sensors to perceive their surroundings and make informed, autonomous decisions. The security of these systems has become increasingly critical, as malicious actors can exploit vulnerabilities in the perception pipeline, leading to potentially catastrophic consequences. Our research focuses on studying the security vulnerabilities of sensor perception and decision-making modules in autonomous systems, including their sensing mechanisms along with the ML-based object detection, tracking and planning algorithms. For example, an adversary can remotely inject deceptive patterns into camera feeds, creating or altering objects in the perceived environment, causing unsafe control actions. To counter such threats, we will introduce a novel defense framework that leverages spatiotemporal consistency checks, which is agnostic to the specific sensing modality or attack vector.

Connected and Autonomous Vehicles (CAVs) will revolutionize transportation, promising enhanced safety and efficiency. Vehicle-to-Everything (V2X) connectivity enables vehicles and infrastructure to share information, fostering cooperative perception (CP) and decision making, enabling groundbreaking applications like cooperative driving, dynamic map updates, platooning, and infrastructure-assisted traffic management. Safety-critical CAV applications demand stringent performance requirements, including high perception accuracy, low end-to-end latency, and high reliability. Edge-assisted CP systems face challenges in scalable raw sensor data sharing from multiple vehicles and adapting to dynamic network conditions. Our research addresses these challenges by proposing a goal-oriented (semantic) communications framework, which leverages ML techniques to intelligently process and extract the most relevant information from the sensor data, and jointly optimizes the computation and communication resources at the vehicles and edge to meet the end-goals of CP. This research has broader implications, extending beyond CAVs to various multi-agent systems.