University of Nevada, Reno
Privacy and Security for Trustworthy AI/ML in Multi-Agent Critical Infrastructures: An Analysis of Adversarial Dynamics and Protective Strategies
Abstract
dc:description.abstractThis dissertation analyzes the pressing security and privacy challenges in Multi-Agent Critical Infrastructures ( MACIs ), which increasingly integrate cyber-physical system components like IoT devices, sensors, UAVs, and cloud technologies. These components are prolific data generators and are crucial for operational efficacy and enhanced decision-making across various artificial intelligence (AI) and machine learning (ML)-driven MACI sectors. However, the integration of state-of-the-art AI/ML techniques in MACIs, although beneficial for predictive analytics and decision support, introduces significant vulnerabilities to sophisticated cyberattacks. The raw MACI data-serving as the training data for these AI/ML frameworks - often contain sensitive user and operational information, necessitating robust privacy and security measures to protect against unauthorized access and manipulation. Thus, there is a growing need for robust data privacy and security methods to be integrated into the information-sharing and learning process of these frameworks in order to preserve data confidentiality and authenticity. In this context, this dissertation analyzes the inherent vulnerabilities and threats to existing information exchange and techniques as well as their exploitation and impacts on the disclosure of private information and MACI operations. Central to our analysis is the scrutiny of differential privacy (DP), a prevalent method for safeguarding data privacy, which we find is susceptible to adversarial exploitation. Our analysis reveals that DP paradoxically introduces exploitable loopholes that adversaries can leverage to conduct stealthy attacks such as false data injection and poisoning in various learning frameworks including federated learning ( FL ) and cooperative multi-agent reinforcement learning ( CMARL ). To address these threats, the dissertation presents a comprehensive theoretical analysis from an adversarial perspective, developing a novel analytical framework that establishes a provable correlation between DP parameters and adversarial objectives to aid in designing robust, fault-tolerant systems. Particularly, we present novel adversarial strategies like differential privacy-exploited stealthy model poisoning (DeSMP) attacks in FL. Further investigation in this direction led us to develop an adaptive model poisoning through episodic loss memorization (α-MPELM) technique, which deceives state-of-the-art Norm , Accuracy , and Mix ( Norm + Accuracy ) anomaly detection algorithms by 6.8%, 12.6%, and 13.8%, respectively. In response to these challenges, we introduce a reinforcement learning (RL)-based DP level selection strategy, termed rDP, which dynamically adjusts the necessary level of DP noise - small enough to keep the poisoning window tolerable; yet large enough to maintain a satisfactory level of data privacy. Addressing localized poisoning in CMARL, the dissertation proposes the privacy-exploited local poisoning attack (PeLPA), which exploits DP noise to hinder model convergence effectively, increasing steps to goal by up to 64.41% and prolonging convergence times by 1.38x. As a countermeasure, we developed reinforcing autonomous multi-agent protection through adversarial resistance in transportation (RAMPART), a generative adversarial network (GAN)-based defense mechanism that filters malicious data and generates synthetic samples to maintain model integrity, achieving an F1 score of 0.852 and an accuracy of 96.3%. Moreover, we analyze adversarial manipulation in CMARL's experience-sharing processes, where we introduce the Byzantine robust neighbor experience-sharing (BRNES) framework to mitigate Byzantine attacks through dynamic neighbor selection and weighted experience aggregation. BRNES is faster in response times by up to 8.32x compared to non-private frameworks. This comprehensive analysis of the aforementioned adversarial tactics and the development of several novel defensive measures in this dissertation will enable future researchers and developers to design, deploy, verify, and maintain trustworthy AI/ML frameworks for next-generation MACI operations. Additionally, it will benefit them to conduct further research and development to mitigate unprecedented and time-constrained adversarial techniques through neuroadaptive modeling, continual learning, and secured neurocognitive information exchange.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- Doctorate Degree
- Department thesis:degree_department
- Computer Science
- Grantor
- University of Nevada, Reno
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hossain, Md. Tamjid
- Advisor dc:contributor.advisor
-
- La, Hung Manh
- Committee members dc:contributor.committeemember
-
- Badsha, Shahriar
- Zhao, Dongfang
- Nguyen, Tin
- Ben-Idris, Mohammed
- Xu, Hao
Subjects
dc:subject × 6Rights
dc:rights- Statement dc:rights
-
- Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
- Author(s)
Identifiers
dc:identifier.*- Repository record dc:identifier.uri
- https://scholarwolf.unr.edu/handle/11714/10976