Georgia Institute of Technology
Perfectly Undetectable False Data Injection Attacks to Robotics with Probabilistic Defense Mechanism
Abstract
dc:description.abstractAdversarial attacks against Cyber-Physical Systems (CPS) are a pressing concern in the ever-increasingly connected world. Proper security of such systems requires a threefold guarantee on system data: confidentiality, integrity, and availability. These guarantees pose a significant challenge to the reliable deployment of CPS. In particular, the False Data Injection Attack (FDIA) is an injection of false data to communication lines, causing a loss of data integrity and undesired control system responses. This research seeks to reveal a new potential of undetectable FDIA on CPS, particularly on networked robotic systems, and provides a probabilistic defense framework for such attacks while simultaneously maintaining data confidentiality. Aim 1's objective is to create a perfectly undetectable, affine transformation-based FDIA, which operates with full undetectability regardless of controller or attack detector. In aim 1, we ask: Can an attacker design observable and control command affine transformations to realize perfect undetectability for a Single Integrator (SI) system? To accomplish this task, rules for coordinated affine transformations on communication lines are defined for an SI modeled robotic manipulator, coined the Affine Transformation Perfectly-undetectable Attack (ATPA). The ATPA has an advantage over covert attacks because it does not rely on a real-time, simulated, fake plant, but rather static signal transformations. These conditions are verified on a FANUC 6 DoF manipulator with an adaptive attack detector. Experiments find the ATPA to be perfectly undetectable, completely bypassing the attack detector. Aim 2 seeks to extend the ATPA to multi-agent consensus-controlled systems and to determine where model simplifications by the attacker and/or defender are detrimental to their individual objectives. In aim 2, we ask: Can the attacker design observable and control command affine transformations to realize perfect undetectability for time variant SI and unicycle modeled, consensus controlled Multi-Agent Systems (MAS)? Additionally, we ask which modeling decisions by the attacker and defender increase and/or decrease the probability of attack detection? Two methods of consensus control ATPA are formed, the Individual Agent Dynamics Matching (IADM), similar to the ATPA formed in Aim 1, and Close Loop Matching (CLM). A fundamental contribution is the CLM ATPA which exploits the uncontrollable modes inherent in consensus formation control, creating new attack vulnerabilities absent in single-agent systems. These vulnerabilities allow the attacker to alter attack parameters while maintaining the same objective, resulting in more flexible ATPA. Furthermore, often in the literature, attackers and/or defenders estimate unicycle agents as SI models, resulting in minor errors in execution. We evaluate these cases to determine where defender-side and attacker-side model simplifications cause consensus and trajectory errors. Simulations verify theoretical findings, and experiments are executed on the Robotarium multi-robot platform where hardware validity is confirmed. When the attacker and defender each model their agents as unicycle models instead using the SI model simplification, their individual objectives are insignificantly effected, and the attacked observables remain within expected error bounds. Aim 3 seeks to build a defense mechanism for detecting both detectable and undetectable FDIA while maintaining data confidentiality. In aim 3, we ask: Can the defender utilize an encrypted framework to simultaneously defend against eavesdropping, detectable, and undetectable FDIA? Homomorphic Encryption (HE) allows networked control system signals and parameters to remain encrypted while operating in the cloud, but HE signals are still vulnerable to malleability-based FDIA. The derived defense mechanism uses a Modified Model Reference Adaptive Controller (MMRAC) algorithm which allows for HE without noise overflow issues while mitigating smaller FDIA by preserving the adaptive component of the algorithm. An encryption noise-based overflow trap is designed to prevent larger FDIA that would otherwise throw the system into instability due to encryption noise overflow. A defense architecture against undetectable attacks is formed through the fusion of two complementary detection protocols, a State Monitoring Signature Function (SMSF) and the overflow trap, working in conjunction with the HE. The mechanism's strengths and weaknesses are analyzed for the two forms of undetectable FDIA: the ATPA and covert attack. Simulations validate the confidentiality and adaptability of MMRAC in the cypherspace. The undetectable FDIA and defense framework are implemented on a physical, mobile unicycle robot, and performance is analyzed, resulting in a robust defense mechanism against undetectable FDIA. The contributions of these aims provide networked robotics with a further understanding of adversary FDIA potential and how to defend such systems using a carefully crafted defense mechanism.
Degree
thesis:*- Name thesis:degree_name
- Mechanical Engineering, PhD
- Department thesis:degree_department
- Mechanical Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Blevins, Jacob
- Advisor dc:contributor.advisor
-
- Ueda, Jun
- Committee members dc:contributor.committeemember
-
- Kogiso, Kiminao
- Kousik, Shreyas
- Zhao, Ye
- Coogan, Samuel
- ORCID person:identifier.orcid
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82151