ADAPT: A Game-Theoretic and Neuro-Symbolic Framework for Automated Distributed Adaptive Penetration Testing

TL;DR

This work presents ADAPT, a game-theoretic and neuro-symbolic framework for automated distributed adaptive penetration testing, specifically designed to address the unique cybersecurity challenges of AI-enabled healthcare infrastructure networks.

Abstract

The integration of AI into modern critical infrastructure systems, such as healthcare, has introduced new vulnerabilities that can significantly impact workflow, efficiency, and safety. Additionally, the increased connectivity has made traditional human-driven penetration testing insufficient for assessing risks and developing remediation strategies. Consequently, there is a pressing need for a distributed, adaptive, and efficient automated penetration testing framework that not only identifies vulnerabilities but also provides countermeasures to enhance security posture. This work presents ADAPT, a game-theoretic and neuro-symbolic framework for automated distributed adaptive penetration testing, specifically designed to address the unique cybersecurity challenges of AI-enabled healthcare infrastructure networks. We use a healthcare system case study to illustrate the methodologies within ADAPT. The proposed solution enables a learning-based risk assessment. Numerical experiments are used to demonstrate effective countermeasures against various tactical techniques employed by adversarial AI.

Figures (6)

• Figure 1: The framework of the ADAPT: The upper half illustrates an online automated adaptation of penetration testing. It integrates game-theoretic and neuro-symbolic frameworks, consisting of five distinct building blocks (to be introduced in Section III). The lower half depicts an example of AI-enabled healthcare infrastructure. This AI-enabled infrastructure presents an expanded attack surface due to interconnectivity and zero-day vulnerabilities.
• Figure 2: The flow diagram of the symbolic adaptation. Attack models are represented by Multi-Type Graphs (MTGs), and an exploration is performed on this model. When new properties or vulnerabilities in the model are discovered, the model is updated through knowledge searching.
• Figure 3: An illustration of a hospital network with AI data center, consisting of four different areas (Internet Cloud, workstations, access layer, and data center) that represent the transition from the public network to the protected private network of the hospital. The data center includes two different subnets. One is for only data storage, and the other is for both data storage and AI applications. The attacker will access the network from the web server in workstations and target the AI center as the network's critical asset.
• Figure 4: An illustration of the meta-security game for in the hospital network. For simplification, only the path to the critical asset, i.e., the AI center, is considered. The local penetration plans in the micro games influence the global attack strategy, while the policy evaluation in the macro process helps provide the utilities for the micro games.
• Figure 5: An illustration of MTGs for different impact techniques. Red nodes represent the attacker, blue nodes represent the defender, yellow nodes represent the system randomness, and green nodes represent outcomes.

Theorems & Definitions (4)

• Definition 1: Meta-Security Game
• Definition 2: Meta Penetration Playbook
• Definition 3: Network Risk Score
• Definition 4: Knowledge Library