ADAPT: A Game-Theoretic and Neuro-Symbolic Framework for Automated Distributed Adaptive Penetration Testing thanks: 2Authors contributed equally, 1Corresponding author, Authors belong to Department of Electrical and Computer Engineering New York University, New York 11201, USA{hl4155, yg2047, qz494}@nyu.edu

Haozhe Lei12, Yunfei Ge 2, Quanyan Zhu
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.

I Introduction

Modern artificial intelligence (AI), such as machine learning (ML) technologies, are becoming increasingly integrated into many infrastructures, including smart transportation systems and healthcare infrastructures. In healthcare, they have shown the potential to help healthcare infrastructure in patient scheduling [1, 2], pathological analysis [3], and care management [4]. While there are significant benefits, there are concerns regarding zero-day vulnerabilities and the expanded attack surface.

Penetration testing is a valuable ethical hacking method for uncovering vulnerabilities in increasingly complex infrastructures and devising remediation strategies. As these infrastructures become more complex, with millions of interconnected devices, scalability emerges as a critical challenge. It is essential to develop a distributed, modular, and automated approach that addresses device-level testing needs while considering global influences through interconnectivity. Another challenge stems from the dynamic nature of networked devices and their vulnerabilities. There is a growing need for adaptive and automated approaches to continuously update the vulnerability landscape, ensuring that threats are exhaustively identified, risks accurately assessed, and remediation measures properly applied. The third challenge arises from the integration of AI capabilities into the infrastructure. The emergence of adversarial AI/ML introduces new and evolving threat vectors, which are designed to evade detection and testing. There is a need for the development of automated and strategic approaches that can intelligently outmaneuver their evolving nature through continuous knowledge acquisition and learning. To this end, we establish a game-theoretic and neuro-symbolic framework for automated distributed adaptive penetration testing (ADAPT).

Refer to caption
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.

To address the challenges in penetration testing within AI-enabled healthcare infrastructure networks, ADAPT consolidates the game-theoretic framework and the neuro-symbolic framework, shown in the Figure 1. The game-theoretic and neuro-symbolic framework consists of five building blocks. The macro-game and micro-game blocks serve as representations of the given system. Game-theoretic strategies are updated based on the selected attack models through neural learning in the online learning blocks. Different attack models are represented using game trees, which encode relevant attack and defense actions selected from the knowledge. This knowledge contains vulnerabilities of the health infrastructure that are shared among multiple stakeholders in the medical system. When new paths or vulnerabilities are discovered through exploration (e.g., automated fuzzing techniques [5]) and penetration testing, the knowledge base is generated and updated accordingly. ADAPT helps medical systems evaluate their AI-enabled network for scalability, the impact of reachability, and the exhaustiveness of risk identification, and protects the confidentiality, integrity, and availability of the AI model. The purpose of it is to ensure preparedness against continuously evolving malicious threats such as ransomware and zero-day attacks on healthcare infrastructures.

II Related Works

Traditional manual penetration testing performed by skilled IT professionals is time-consuming, resource-intensive, and prone to human error [6]. Current automated penetration testing methods, despite advancements in efficiency, are becoming increasingly non-standard, complex, and resource-intensive. Reinforcement learning (RL) or Markov Decision Process (MDP) based methods [7, 8] suffer from the curse of dimensionality, as they define the state space as the collection of all known information for each machine on the network. Partially Observable Markov Decision Process (POMDP) methods [9] face scalability issues, making it unfeasible to model and solve for large networks.

As state-of-the-art technology, AI models present relatively higher risks compared to traditional methods [10]. Several studies have demonstrated that AI models can be easily subjected to malicious data-poisoning attacks in not only regression tasks [11] but also image classification [12] and robotics control [13]. In addition to directly degrading the ML model, evidence also points to various other types of attacks, such as DDoS attacks [14] and bypassing attacks [15]. These findings underscore the significant threat posed by attackers and a wide range of vulnerabilities for the AI-enabled system.

III ADAPT Framework and Methodologies

This section, we present the ADAPT framework introduced in Figure 1. It consists of a meta-game framework and a neuro-symbolic framework.

III-A Meta-game-based Automated Penetration Testing

We propose a meta-security game over a network graph where the macro strategic game represents strategic attack activities between nodes, while the micro tactic game details tactic-level attack procedures on each local node. Let the directed graph 𝒢=𝒱,𝒢𝒱\mathcal{G}=\langle\mathcal{V},\mathcal{E}\ranglecaligraphic_G = ⟨ caligraphic_V , caligraphic_E ⟩ represent the target network topology, where 𝒱𝒱\mathcal{V}caligraphic_V is a set of nodes (e.g., server, database, device), and 𝒱×𝒱𝒱𝒱\mathcal{E}\subseteq\mathcal{V}\times\mathcal{V}caligraphic_E ⊆ caligraphic_V × caligraphic_V is a set of directed edges representing connections (e.g., SSH, RDP, cloud services) from node u𝑢uitalic_u to node v𝑣vitalic_v. Self-loops are allowed as they indicate continued exploration of the same node. Let v0𝒱superscript𝑣0𝒱v^{0}\in\mathcal{V}italic_v start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT ∈ caligraphic_V be the initial foothold in the system. As an ethical attacker, the penetration tester aims to explore available information, exploit discovered vulnerabilities, and influence critical assets inside the healthcare infrastructure network.

Micro Tactic Games: To model the interactions between the attacker and the defender at each local node, we use extensive-form game tree to explicitly and visually represent the sequential moves, possible outcomes, and information available at each decision point. We assume that all players have perfect recall; i.e., the player remembers every piece of information that he knows from the past, including his moves, the other player’s moves, or chance moves. The structure of the game tree is inspired by the intrusion kill chain [16], a concept that outlines the structure of intrusions. This model guides analysis to inform actionable security intelligence.

For each node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V in the network, the Micro Tactic Game (MTG) is defined as Γv=𝒩{c},v,P,{𝒜iv}i𝒩c,σcv,{uiv}i𝒩,𝒵vsuperscriptΓ𝑣𝒩𝑐superscript𝑣𝑃subscriptsubscriptsuperscript𝒜𝑣𝑖𝑖𝒩𝑐subscriptsuperscript𝜎𝑣𝑐subscriptsubscriptsuperscript𝑢𝑣𝑖𝑖𝒩superscript𝒵𝑣\Gamma^{v}=\langle\mathcal{N}\cup\{c\},\mathcal{H}^{v},P,\{\mathcal{A}^{v}_{i}% \}_{i\in\mathcal{N}\cup c},\sigma^{v}_{c},\{u^{v}_{i}\}_{i\in\mathcal{N}},% \mathcal{Z}^{v}\rangleroman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT = ⟨ caligraphic_N ∪ { italic_c } , caligraphic_H start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT , italic_P , { caligraphic_A start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_i ∈ caligraphic_N ∪ italic_c end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , { italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_i ∈ caligraphic_N end_POSTSUBSCRIPT , caligraphic_Z start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ⟩. Here, 𝒩={a,d}𝒩𝑎𝑑\mathcal{N}=\{a,d\}caligraphic_N = { italic_a , italic_d } represents the players, the attacker (a) and the defender (d), while c𝑐citalic_c represents system randomness with fixed policy σcvsubscriptsuperscript𝜎𝑣𝑐\sigma^{v}_{c}italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT. Each vertex hvsuperscript𝑣h\in\mathcal{H}^{v}italic_h ∈ caligraphic_H start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT in the game tree represents a sequence of actions, referred to as history. The function P(h)𝑃P(h)italic_P ( italic_h ) determines whose turn it is at each decision point (attacker, defender, or nature) for a given history vertex hhitalic_h. 𝒜ivsubscriptsuperscript𝒜𝑣𝑖\mathcal{A}^{v}_{i}caligraphic_A start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT is a set of actions available to player i𝑖iitalic_i, and A(h)𝐴A(h)italic_A ( italic_h ) describes the feasible actions for the player at vertex hhitalic_h. 𝒵vsuperscript𝒵𝑣\mathcal{Z}^{v}caligraphic_Z start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT represents the possible outcomes for each tactic in the game, corresponding to the results at the leaf vertices of the game tree. We assume the outcomes are either remaining in the current node or leading to another connected node, denoted as 𝒵v={uu𝒱,(v,u)}superscript𝒵𝑣conditional-set𝑢formulae-sequence𝑢𝒱𝑣𝑢\mathcal{Z}^{v}=\{u\mid u\in\mathcal{V},(v,u)\in\mathcal{E}\}caligraphic_Z start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT = { italic_u ∣ italic_u ∈ caligraphic_V , ( italic_v , italic_u ) ∈ caligraphic_E }. Finally, the utility function uivsubscriptsuperscript𝑢𝑣𝑖u^{v}_{i}italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT determines the payoff or cost for player i𝑖iitalic_i when reaching a certain outcome.

By solving the extensive-form game, we are able to obtain the optimal local penetration for the attacker under possible defense plans. Given the nature’s fixed policy (if any) and the plan profile of the attacker and the defender, i.e., Φv=(σav,σdv,σcv)superscriptΦ𝑣subscriptsuperscript𝜎𝑣𝑎subscriptsuperscript𝜎𝑣𝑑subscriptsuperscript𝜎𝑣𝑐\Phi^{v}=(\sigma^{v}_{a},\sigma^{v}_{d},\sigma^{v}_{c})roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT = ( italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ), we denote τv(zΦv)[0,1]superscript𝜏𝑣conditional𝑧superscriptΦ𝑣01\tau^{v}(z\mid\Phi^{v})\in[0,1]italic_τ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ( italic_z ∣ roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ) ∈ [ 0 , 1 ] as the tactic outcome probability, which is the joint probability of reaching that outcome under ΦvsuperscriptΦ𝑣\Phi^{v}roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT.

Macro Strategic Process: One key component in the MTG is the utility function for each outcome, uiv(z)superscriptsubscript𝑢𝑖𝑣𝑧u_{i}^{v}(z)italic_u start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ( italic_z ), for all z𝒵v𝑧superscript𝒵𝑣z\in\mathcal{Z}^{v}italic_z ∈ caligraphic_Z start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT and i𝒩𝑖𝒩i\in\mathcal{N}italic_i ∈ caligraphic_N. Utilities represent the payoff or cost of staying or moving to the next node and must be evaluated globally, considering neighboring nodes and their connections. After local exploration and exploitation, the attacker can use obtained credentials or discovered vulnerabilities to move to different nodes, a process known as lateral movement. The attacker’s movement and the creation of the attack kill chain depend on the network topology and the expected utilities of each node. We model this decision-making process across the network using an MDP, a Macro Strategic Process (MSP).

The Macro Strategic Process (MSP) for the attacker within the MEGA-PT framework is characterized by a tuple Λg=𝒮,𝒜g,T,R,γsuperscriptΛ𝑔𝒮superscript𝒜𝑔𝑇𝑅𝛾\Lambda^{g}=\langle\mathcal{S},\mathcal{A}^{g},T,R,\gamma\rangleroman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT = ⟨ caligraphic_S , caligraphic_A start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT , italic_T , italic_R , italic_γ ⟩, where 𝒮=𝒱𝒮𝒱\mathcal{S}=\mathcal{V}caligraphic_S = caligraphic_V represents the network nodes as states, 𝒜g=superscript𝒜𝑔\mathcal{A}^{g}=\mathcal{E}caligraphic_A start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT = caligraphic_E represents the connections between nodes as the action space, T𝑇Titalic_T denotes the transition success probability function, R𝑅Ritalic_R represents the movement rewards, and γ𝛾\gammaitalic_γ is the discounting factor. The success of an attack attempt is influenced by the attacker’s capability. For the purposes of this analysis, the transition success probability is defined as follows: if the attacker opts to remain at the same node, the probability of staying in the same state is 1111. When the attacker chooses to move to a different node via an outgoing edge, the probability of a successful transition to the new node is denoted by ca[0,1]subscript𝑐𝑎01c_{a}\in[0,1]italic_c start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ∈ [ 0 , 1 ], reflecting the attacker’s skill level. If the attack attempt to move is unsuccessful, the attacker remains in the current node. The reward function R𝑅Ritalic_R provides positive rewards V¯(v)¯𝑉superscript𝑣\bar{V}(v^{\prime})over¯ start_ARG italic_V end_ARG ( italic_v start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) based on importance of the node for moving to the next node v𝒱superscript𝑣𝒱v^{\prime}\in\mathcal{V}italic_v start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_V and negative penalties Masubscript𝑀𝑎superscriptM_{a}\in\mathds{R}^{-}italic_M start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT for remaining at the same node without progress.

Meta-Security Game: Unlike traditional MDPs, where the attacker can freely choose actions to optimize expected utility, in the realistic penetration testing settings, the attack strategy in the network-level depends on explorations at the local nodes. If the attacker does not find any vulnerabilities leading to the next node, they cannot move forward. Therefore, the global attacks strategy in the MSP relies on the outcomes of the MTG. For each node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V, the optimal local penetration plans determine the probability of each tactic outcome, denoted as τv(z)superscript𝜏𝑣𝑧\tau^{v}(z)italic_τ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ( italic_z ), where z𝑧zitalic_z represents an outgoing edge from v𝑣vitalic_v. This probability indicates the likelihood that the attacker will select a particular action ag=(v,z)superscript𝑎𝑔𝑣𝑧a^{g}=(v,z)italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT = ( italic_v , italic_z ) in the global attack strategy. For node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V, given the MTG ΓvsuperscriptΓ𝑣\Gamma^{v}roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT and the local plan profile ΦvsuperscriptΦ𝑣\Phi^{v}roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT, the global attack strategy is given by

πg(ag=(v,z)s=v)=τv(zΦv),z𝒵v.formulae-sequencesuperscript𝜋𝑔superscript𝑎𝑔conditional𝑣𝑧𝑠𝑣superscript𝜏𝑣conditional𝑧superscriptΦ𝑣for-all𝑧subscript𝒵𝑣\displaystyle\pi^{g}(a^{g}=(v,z)\mid s=v)=\tau^{v}(z\mid\Phi^{v}),\qquad% \forall z\in\mathcal{Z}_{v}.italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ( italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT = ( italic_v , italic_z ) ∣ italic_s = italic_v ) = italic_τ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ( italic_z ∣ roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ) , ∀ italic_z ∈ caligraphic_Z start_POSTSUBSCRIPT italic_v end_POSTSUBSCRIPT . (1)

Policy evaluation estimates the effectiveness of the global attack strategy, πgsuperscript𝜋𝑔\pi^{g}italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT, by calculating expected cumulative utilities. This involves computing value functions using Bellman equations. For each state s𝒮𝑠𝒮s\in\mathcal{S}italic_s ∈ caligraphic_S, the value function under πgsuperscript𝜋𝑔\pi^{g}italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT is given by:

V𝑉\displaystyle Vitalic_V (s)πg={}^{\pi^{g}}(s)=start_FLOATSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_FLOATSUPERSCRIPT ( italic_s ) = (2)
ag𝒜gπg(ags)s𝒱T(ss,ag)[R(s,ag,s)+γVπg(s)].subscriptsuperscript𝑎𝑔superscript𝒜𝑔superscript𝜋𝑔conditionalsuperscript𝑎𝑔𝑠subscriptsuperscript𝑠𝒱𝑇conditionalsuperscript𝑠𝑠superscript𝑎𝑔delimited-[]𝑅𝑠superscript𝑎𝑔superscript𝑠𝛾superscript𝑉superscript𝜋𝑔superscript𝑠\displaystyle\sum_{a^{g}\in\mathcal{A}^{g}}\pi^{g}(a^{g}\mid s)\sum_{s^{\prime% }\in\mathcal{V}}T(s^{\prime}\mid s,a^{g})\left[R(s,a^{g},s^{\prime})+\gamma V^% {\pi^{g}}(s^{\prime})\right].∑ start_POSTSUBSCRIPT italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ∈ caligraphic_A start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUBSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ( italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ∣ italic_s ) ∑ start_POSTSUBSCRIPT italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_V end_POSTSUBSCRIPT italic_T ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∣ italic_s , italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ) [ italic_R ( italic_s , italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT , italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) + italic_γ italic_V start_POSTSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) ] .

This value represents the expected return starting from state s𝑠sitalic_s and following the global attack strategy πgsuperscript𝜋𝑔\pi^{g}italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT. In the MTG, the utility of each outcome z𝑧zitalic_z at node v𝑣vitalic_v reflects the expected reward from taking that action and moving to the next node. The utility functions in the MTG are defined as uav(z=u)=s𝒱T(ss=v,ag=(v,u))[R(s,ag,s)+γVπg(s)].subscriptsuperscript𝑢𝑣𝑎𝑧𝑢subscriptsuperscript𝑠𝒱𝑇formulae-sequenceconditionalsuperscript𝑠𝑠𝑣superscript𝑎𝑔𝑣𝑢delimited-[]𝑅𝑠superscript𝑎𝑔superscript𝑠𝛾superscript𝑉superscript𝜋𝑔superscript𝑠u^{v}_{a}(z=u)=\sum_{s^{\prime}\in\mathcal{V}}T(s^{\prime}\mid s=v,a^{g}=(v,u)% )\left[R(s,a^{g},s^{\prime})+\gamma V^{\pi^{g}}(s^{\prime})\right].italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ( italic_z = italic_u ) = ∑ start_POSTSUBSCRIPT italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_V end_POSTSUBSCRIPT italic_T ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∣ italic_s = italic_v , italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT = ( italic_v , italic_u ) ) [ italic_R ( italic_s , italic_a start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT , italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) + italic_γ italic_V start_POSTSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) ] . The defender’s utility is the negative of the attacker’s utility: udv(z)=uav(z)subscriptsuperscript𝑢𝑣𝑑𝑧subscriptsuperscript𝑢𝑣𝑎𝑧u^{v}_{d}(z)=-u^{v}_{a}(z)italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT ( italic_z ) = - italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ( italic_z ) for all z𝒵v𝑧superscript𝒵𝑣z\in\mathcal{Z}^{v}italic_z ∈ caligraphic_Z start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT.

The exploration at each local MTG generates the global attack strategy, while the estimated value for each node through policy evaluation under the current strategy represent the expected outcome utilities at each MTG. Together, the MSP and the MTGs constitute a meta-security game that captures decision-making in penetration testing at both network and node levels. A detailed example can be found in Section IV.

Definition 1 (Meta-Security Game).

Given the network system graph 𝒢=𝒱,𝒢𝒱\mathcal{G}=\langle\mathcal{V},\mathcal{E}\ranglecaligraphic_G = ⟨ caligraphic_V , caligraphic_E ⟩, the meta-security game is composed of two parts: Ξ={Γv}v𝒱,ΛgΞsubscriptsuperscriptΓ𝑣𝑣𝒱superscriptΛ𝑔\Xi=\langle\{\Gamma^{v}\}_{v\in\mathcal{V}},\Lambda^{g}\rangleroman_Ξ = ⟨ { roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩, where {Γv}v𝒱subscriptsuperscriptΓ𝑣𝑣𝒱\{\Gamma^{v}\}_{v\in\mathcal{V}}{ roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT is the set of MTGs and ΛgsuperscriptΛ𝑔\Lambda^{g}roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT is the MSP.

The MSP and the MTGs are inherently coupled, hence, a holistic solution concept is necessary for the proposed meta-security game.

Definition 2 (Meta Penetration Playbook).

Consider the meta-security game Ξ={Γv}v𝒱,ΛgΞsubscriptsuperscriptΓ𝑣𝑣𝒱superscriptΛ𝑔\Xi=\langle\{\Gamma^{v}\}_{v\in\mathcal{V}},\Lambda^{g}\rangleroman_Ξ = ⟨ { roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩ defined in Definition 1, the meta penetration playbook ξ={Φv}v𝒱,πg𝜉subscriptsuperscriptΦ𝑣𝑣𝒱superscript𝜋𝑔\xi=\langle\{\Phi^{v}\}_{v\in\mathcal{V}},\pi^{g}\rangleitalic_ξ = ⟨ { roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩ is composed of the local penetration profiles at each node and the global attack strategy. They satisfy two conditions: 1) policy Dependency - the global attack strategy πasuperscript𝜋𝑎\pi^{a}italic_π start_POSTSUPERSCRIPT italic_a end_POSTSUPERSCRIPT at the macro strategy process depends on the local penetration plans {Φv}v𝒱subscriptsuperscriptΦ𝑣𝑣𝒱\{\Phi^{v}\}_{v\in\mathcal{V}}{ roman_Φ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT; 2) value Dependency - for each MTG at node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V, the utility of each tactic expected outcome depends on the policy evaluation results of global attack strategy πgsuperscript𝜋𝑔\pi^{g}italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT

Definition 3 (Network Risk Score).

Consider the meta-security game ΞΞ\Xiroman_Ξ defined in Definition 1 and the corresponding meta penetration playbook ξ𝜉\xiitalic_ξ defined in Definition 2, the network risk score of node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V is defined as ω(vξ)={Vπg(v)Vmaxif Vπg(v)0,0otherwise,𝜔conditional𝑣𝜉casessuperscript𝑉superscript𝜋𝑔𝑣subscript𝑉𝑚𝑎𝑥if superscript𝑉superscript𝜋𝑔𝑣00otherwise,\omega(v\mid\xi)\!\!=\!\!\begin{cases}\frac{V^{\pi^{g}}(v)}{V_{max}}&\!\text{% if }V^{\pi^{g}}(v)\!\geq\!0,\\ 0&\!\text{otherwise,}\end{cases}italic_ω ( italic_v ∣ italic_ξ ) = { start_ROW start_CELL divide start_ARG italic_V start_POSTSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ( italic_v ) end_ARG start_ARG italic_V start_POSTSUBSCRIPT italic_m italic_a italic_x end_POSTSUBSCRIPT end_ARG end_CELL start_CELL if italic_V start_POSTSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ( italic_v ) ≥ 0 , end_CELL end_ROW start_ROW start_CELL 0 end_CELL start_CELL otherwise, end_CELL end_ROW where Vπg(v)superscript𝑉superscript𝜋𝑔𝑣V^{\pi^{g}}(v)italic_V start_POSTSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ( italic_v ) is the policy evaluation value function under the meta penetration playbook.

III-B Neural-Symbolic Penetration Algorithm

Consider a partially known meta-security game Ξ={Γv}v𝒱,ΛgΞsubscriptsuperscriptΓ𝑣𝑣𝒱superscriptΛ𝑔\Xi=\langle\{\Gamma^{v}\}_{v\in\mathcal{V}},\Lambda^{g}\rangleroman_Ξ = ⟨ { roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩, assuming that unknown information only occurs at the entry point and the edge node of the network. From the attacker’s perspective, denote these nodes as the web server vosubscript𝑣𝑜v_{o}italic_v start_POSTSUBSCRIPT italic_o end_POSTSUBSCRIPT and the AI center server vdsubscript𝑣𝑑v_{d}italic_v start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT that stores and trains ML models. To solve this meta-game with incomplete information, we present the algorithm for solving a meta-penetration testing playbook, followed by an explanation of how neural-symbolic libraries assist in automating adaptation to networks with incomplete information.

Refer to caption
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.

III-B1 Computation of the Meta Penetration Playbook

To describe the solution of our framework, we assume that the meta-security game ΞΞ\Xiroman_Ξ operates within a network with complete information. To determine the optimal meta-penetration playbook ξ𝜉\xiitalic_ξ for ΞΞ\Xiroman_Ξ under the chosen solution concept, we propose an iterative computation algorithm in Algorithm 1. For illustration purposes, we assume that in each MTG, players aim to find the Nash equilibrium as their micro penetration profile Φv,eq=(σav,eq,σdv,eq,σcv)superscriptΦ𝑣𝑒𝑞subscriptsuperscript𝜎𝑣𝑒𝑞𝑎subscriptsuperscript𝜎𝑣𝑒𝑞𝑑subscriptsuperscript𝜎𝑣𝑐\Phi^{v,eq}=(\sigma^{v,eq}_{a},\sigma^{v,eq}_{d},\sigma^{v}_{c})roman_Φ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT = ( italic_σ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ).

Algorithm 1 Meta Penetration Playbook Computation
1:Input Meta-security game Ξ={Γv}v𝒱,ΛgΞsubscriptsuperscriptΓ𝑣𝑣𝒱superscriptΛ𝑔\Xi=\langle\{\Gamma^{v}\}_{v\in\mathcal{V}},\Lambda^{g}\rangleroman_Ξ = ⟨ { roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩
2:Initialize Arbitrary global attack strategy πgsuperscript𝜋𝑔\pi^{g}italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT
3:while Meta penetration playbook not converge do
4:     Obtain the value function Vπgsuperscript𝑉superscript𝜋𝑔V^{\pi^{g}}italic_V start_POSTSUPERSCRIPT italic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT of ΛgsuperscriptΛ𝑔\Lambda^{g}roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT by (2)
5:     Update the uivsubscriptsuperscript𝑢𝑣𝑖u^{v}_{i}italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT in each MTG ΓvsuperscriptΓ𝑣\Gamma^{v}roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT
6:     For all v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V compute Φv,eq=(σav,eq,σdv,eq,σcv)superscriptΦ𝑣𝑒𝑞subscriptsuperscript𝜎𝑣𝑒𝑞𝑎subscriptsuperscript𝜎𝑣𝑒𝑞𝑑subscriptsuperscript𝜎𝑣𝑐\Phi^{v,eq}=(\sigma^{v,eq}_{a},\sigma^{v,eq}_{d},\sigma^{v}_{c})roman_Φ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT = ( italic_σ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) 
7:     Update πasuperscript𝜋𝑎\pi^{a}italic_π start_POSTSUPERSCRIPT italic_a end_POSTSUPERSCRIPT under {Φv,eq}v𝒱subscriptsuperscriptΦ𝑣𝑒𝑞𝑣𝒱\{\Phi^{v,eq}\}_{v\in\mathcal{V}}{ roman_Φ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT using (1)
8:Return Meta penetration playbook ξ={Φv,eq}v𝒱,\xi=\langle\{\Phi^{v,eq}\}_{v\in\mathcal{V}},italic_ξ = ⟨ { roman_Φ start_POSTSUPERSCRIPT italic_v , italic_e italic_q end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , πg\pi^{g}\rangleitalic_π start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩

For the extensive form game tree design, we can use backward induction to find the subgame perfect Nash equilibrium. Alternatively, a neural network-based approach can be employed to find the game solution if the MTG is computationally complicated.

III-B2 Symbolic Adaptation of MTGs

In the meta-security game with incomplete information, our goal is to explore the MTGs with partially or completely unknown information, and the flow diagram is shown in Figure 2. As an example, we consider the MTGs of the entry node and the edge node in the network, denoted as ΓvosuperscriptΓsubscript𝑣𝑜\Gamma^{v_{o}}roman_Γ start_POSTSUPERSCRIPT italic_v start_POSTSUBSCRIPT italic_o end_POSTSUBSCRIPT end_POSTSUPERSCRIPT and ΓvdsuperscriptΓsubscript𝑣𝑑\Gamma^{v_{d}}roman_Γ start_POSTSUPERSCRIPT italic_v start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT end_POSTSUPERSCRIPT, respectively. We use First-Order-Logic (FOL) representation κvsuperscript𝜅𝑣\kappa^{v}italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT [17] to describe the properties of the network servers in Table I.

      MTG Properties       FOL Representation
      Windows OS       is-windows(v)is-windows𝑣\text{is-windows}(v)is-windows ( italic_v )
      Microsoft protected       is-microsoftD(v)is-microsoftD𝑣\text{is-microsoftD}(v)is-microsoftD ( italic_v )
      Allow Authentication Bypass       can-bypassD(v)can-bypassD𝑣\text{can-bypassD}(v)can-bypassD ( italic_v )
      \cdots       \cdots
TABLE I: Node/Device Properties and Their Corresponding FOL Representations

For each node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V, we assume that there is a finite property set that captures all possible configurations on the node, denoted by 𝒦v:={κlv}l[1,|𝒦v|]assignsuperscript𝒦𝑣subscriptsubscriptsuperscript𝜅𝑣𝑙𝑙1superscript𝒦𝑣\mathcal{K}^{v}:=\{\kappa^{v}_{l}\}_{l\in[1,\cdots|\mathcal{K}^{v}|]}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT := { italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_l ∈ [ 1 , ⋯ | caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT | ] end_POSTSUBSCRIPT. Property exploration refers to the expansion and updating of the property set 𝒦vsuperscript𝒦𝑣\mathcal{K}^{v}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT associated with node v𝑣vitalic_v. In the web server, assuming it is using a Linux operating system and has a Microsoft Defender, the properties are represented as 𝒦vo={is-linux(),is-microsoftD()}superscript𝒦subscript𝑣𝑜is-linuxis-microsoftD\mathcal{K}^{v_{o}}=\{\text{is-linux}(\cdot),\text{is-microsoftD}(\cdot)\}caligraphic_K start_POSTSUPERSCRIPT italic_v start_POSTSUBSCRIPT italic_o end_POSTSUBSCRIPT end_POSTSUPERSCRIPT = { is-linux ( ⋅ ) , is-microsoftD ( ⋅ ) }. Assume the penetration tester is using an automated fuzzing technique [5] to explore the node server and discovers a new vulnerability that allows attacker bypass authentication check. Then, the updated property set of the node is 𝒦vo,={is-linux(),is-microsoftD(),can-bypassD()}superscript𝒦subscript𝑣𝑜is-linuxis-microsoftDcan-bypassD\mathcal{K}^{v_{o},\prime}=\{\text{is-linux}(\cdot),\text{is-microsoftD}(\cdot% ),\text{can-bypassD}(\cdot)\}caligraphic_K start_POSTSUPERSCRIPT italic_v start_POSTSUBSCRIPT italic_o end_POSTSUBSCRIPT , ′ end_POSTSUPERSCRIPT = { is-linux ( ⋅ ) , is-microsoftD ( ⋅ ) , can-bypassD ( ⋅ ) }. Following the same logic, we can describe an AI server using the Windows system with the Windows Defender’s Protection, where an ethical attacker intends to perform an evasion attack on the ML model tactic technique (detail will be described in the following section), as 𝒦vd={is-windows(),is-windowsD(),evadeMLmodel()}superscript𝒦subscript𝑣𝑑is-windowsis-windowsDevadeMLmodel\mathcal{K}^{v_{d}}=\{\text{is-windows}(\cdot),\text{is-windowsD}(\cdot),\text% {evadeMLmodel}(\cdot)\}caligraphic_K start_POSTSUPERSCRIPT italic_v start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT end_POSTSUPERSCRIPT = { is-windows ( ⋅ ) , is-windowsD ( ⋅ ) , evadeMLmodel ( ⋅ ) }.

We use a knowledge library to store all MTGs related to node v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V, formally defined as follows.

Definition 4 (Knowledge Library).

Consider that a knowledge library vsuperscript𝑣\mathcal{L}^{v}caligraphic_L start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT is a set of MTGs corresponding to a specific node v𝑣vitalic_v, defined as v:={Γv(𝒦v)}𝒦vPOW(𝒫v)assignsuperscript𝑣subscriptsuperscriptΓ𝑣superscript𝒦𝑣superscript𝒦𝑣𝑃𝑂𝑊superscript𝒫𝑣\mathcal{L}^{v}:=\{\Gamma^{v}(\mathcal{K}^{v})\}_{\mathcal{K}^{v}\in POW(% \mathcal{P}^{v})}caligraphic_L start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT := { roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ( caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ) } start_POSTSUBSCRIPT caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ∈ italic_P italic_O italic_W ( caligraphic_P start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT, where 𝒫vsuperscript𝒫𝑣\mathcal{P}^{v}caligraphic_P start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT contains all possible node properties for v𝑣vitalic_v and POW(𝒫v)={U|U𝒫v}𝑃𝑂𝑊superscript𝒫𝑣conditional-set𝑈𝑈superscript𝒫𝑣POW(\mathcal{P}^{v})=\{U|U\subseteq\mathcal{P}^{v}\}italic_P italic_O italic_W ( caligraphic_P start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ) = { italic_U | italic_U ⊆ caligraphic_P start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } is the power set of 𝒫vsuperscript𝒫𝑣\mathcal{P}^{v}caligraphic_P start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT.

For given property set 𝒦vsuperscript𝒦𝑣\mathcal{K}^{v}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT and knowledge library vsuperscript𝑣\mathcal{L}^{v}caligraphic_L start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT, a symbolic adaptation process is to find a ΓivvsubscriptsuperscriptΓ𝑣𝑖superscript𝑣\Gamma^{v}_{i}\in\mathcal{L}^{v}roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ∈ caligraphic_L start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT that satisfies:

(l=1|𝒦v|κlv(Γiv))(κv𝒫v𝒦v¬κv(Γiv)) is true.superscriptsubscript𝑙1superscript𝒦𝑣subscriptsuperscript𝜅𝑣𝑙subscriptsuperscriptΓ𝑣𝑖subscriptsuperscript𝜅𝑣superscript𝒫𝑣superscript𝒦𝑣superscript𝜅𝑣subscriptsuperscriptΓ𝑣𝑖 is true.\displaystyle\left(\bigwedge_{l=1}^{|\mathcal{K}^{v}|}\kappa^{v}_{l}(\Gamma^{v% }_{i})\right)\wedge\left(\bigwedge_{\kappa^{v}\in\mathcal{P}^{v}\setminus% \mathcal{K}^{v}}\neg\kappa^{v}(\Gamma^{v}_{i})\right)\text{ is true.}( ⋀ start_POSTSUBSCRIPT italic_l = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT | caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT | end_POSTSUPERSCRIPT italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT ( roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ) ) ∧ ( ⋀ start_POSTSUBSCRIPT italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ∈ caligraphic_P start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ∖ caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ¬ italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ( roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ) ) is true. (3)

Here, we assume that for any possible property set 𝒦vsuperscript𝒦𝑣\mathcal{K}^{v}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT, there exist a feasible node description of v𝑣vitalic_v that satisfies (3). By defining the solution of the penetration playbook as neural and the adaptation of node properties as symbolic, we present a conceptual workflow of ADAPT in Algorithm 2.

Algorithm 2 Neural-Symbolic Penetration Library
1:Input: MSP ΛgsuperscriptΛ𝑔\Lambda^{g}roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT, global graph 𝒢=𝒱,𝒢𝒱\mathcal{G}=\langle\mathcal{V},\mathcal{E}\ranglecaligraphic_G = ⟨ caligraphic_V , caligraphic_E ⟩, knowledge libraries {v}v𝒱subscriptsuperscript𝑣𝑣𝒱\{\mathcal{L}^{v}\}_{v\in\mathcal{V}}{ caligraphic_L start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT, and the initial property set of nodes {𝒦v}v𝒱subscriptsuperscript𝒦𝑣𝑣𝒱\{\mathcal{K}^{v}\}_{v\in\mathcal{V}}{ caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT
2:Symbolic Security Game Adaptation:
3:for v𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V do
4:     while 𝒦vsuperscript𝒦𝑣\mathcal{K}^{v}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT is still evolving do
5:         Explore node v𝑣vitalic_v, find a FOL property κnewvsubscriptsuperscript𝜅𝑣𝑛𝑒𝑤\kappa^{v}_{new}italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n italic_e italic_w end_POSTSUBSCRIPT
6:         if κnewv𝒦vsubscriptsuperscript𝜅𝑣𝑛𝑒𝑤superscript𝒦𝑣\kappa^{v}_{new}\notin\mathcal{K}^{v}italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n italic_e italic_w end_POSTSUBSCRIPT ∉ caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT then
7:              𝒦v=𝒦vκnewvsuperscript𝒦𝑣superscript𝒦𝑣subscriptsuperscript𝜅𝑣𝑛𝑒𝑤\mathcal{K}^{v}=\mathcal{K}^{v}\cup\kappa^{v}_{new}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT = caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT ∪ italic_κ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n italic_e italic_w end_POSTSUBSCRIPT               
8:     Obtain ΓvsuperscriptΓ𝑣\Gamma^{v}roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT by 𝒦vsuperscript𝒦𝑣\mathcal{K}^{v}caligraphic_K start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT with (3)
9:Construct meta-security game Ξ={Γv}v𝒱,ΛgΞsubscriptsuperscriptΓ𝑣𝑣𝒱superscriptΛ𝑔\Xi=\langle\{\Gamma^{v}\}_{v\in\mathcal{V}},\Lambda^{g}\rangleroman_Ξ = ⟨ { roman_Γ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_v ∈ caligraphic_V end_POSTSUBSCRIPT , roman_Λ start_POSTSUPERSCRIPT italic_g end_POSTSUPERSCRIPT ⟩
10:Neural Penetration Playbook Computation:
11:Obtain the meta-penetration playbook ξ𝜉\xiitalic_ξ using Algorithm 1
12:Return the explored meta-security game ΞΞ\Xiroman_Ξ, and the meta penetration playbook ξ𝜉\xiitalic_ξ.

Since the knowledge library may be imperfect and unable to fully describe the node properties, there is a need to constantly update the knowledge library. Due to the limited space, we leave the discussion for future work.

Refer to caption
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.

IV Hospital Case Study

In this section, we discuss a case study of a typical IT infrastructure within a healthcare environment, i.e., a hospital network architecture shown in Figure 3, referring to [18] and [19]. The attacker accesses the hospital network from a web server and targets the AI center in subnet B. Through this example, we aim to highlight the uniqueness of the MTG design in an AI-enhanced network and emphasize the guidance and value of our algorithm for users in such networks.

Real-life examples show the ongoing need for hospitals to deploy AI servers locally [20], typically on the edge device. Consider a hospital deploying an AI model for disease diagnosis using an ML technique similar to those mentioned in the review paper [3]. We assume a malicious attacker (entering the network from the web server node in workstations) aims to steal the access authority of the edge node AI center and perform their attack within the network described in Figure 3. For simplification, the topology of the Meta-security game for this network is considered fixed, except for the edge node AI center, which contains the model development and deployment procedures. The MTG in the edge node AI center is determined by the different impact techniques.

IV-A Penetration Path to access the AI Center

Figure 4 illustrates an example of the meta-security game in the hospital network shown in Figure 3. The attacker starts from the web server and tries to access the AI center by penetrating the AI managerial applications server and the API gateway server. To conserve space in this paper, we present the MTG of the web server following the MITRE ATT&CK framework [21] as an example and omit the others. The MSP defines the global attack strategy by forming an attack kill chain and providing estimated values for each node through policy evaluation under the current strategy. These estimated values represent the expected outcome utilities at each MTG, guiding the formulation of detailed penetration plans at each node. The sequence of attack and defense techniques at the local node influences the global attack strategy from a macro perspective, emphasizing how lateral movement is determined by exploration and exploitation. By continuing the iteration of this process, a meta-solution, i.e., an optimal penetration playbook, can be reached.

Refer to caption
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.

IV-B Impact Techniques

The goal for the attacker is to interrupt, erode confidence in, or destroy the machine learning systems and data. Traditionally, Techniques used for impact can include destroying or tampering with data. Following the MITRE ATLAS framework [22], we use examples to illustrate the differences between three different impact techniques’ MTG trees.

Erode ML Model Integrity: Attacks can happen during the model training, i.e., damage on the model integrity. To influence the ML model’s integrity, there are two primary approaches. The first approach is to directly alter the model, which is efficient and straightforward but also more likely to be detected [23]. The second approach is to poison the training data [11],[13], thereby degrading the model’s performance—often summarized as ”garbage in, garbage out.”

Evade ML Model: Attacks can also occur during model access. An attacker can craft adversarial data that prevents a machine learning model from correctly identifying the contents of the data. This technique can be used to evade downstream tasks where machine learning is utilized. One example of this type of impact technique is the bypassing of Cylance antivirus products [15]. As a result, this impact can cause the adversary’s desired effect on the target model, leading to consequences such as misclassification, missed detections, or maximized energy consumption.

Denial of ML Service: By this impact technique, the attacker is targeting ML systems’ accessibility instead of integrity. The attacker focuses on generating a flood of requests for degrading or shutting down the service. Recall the massive disaster that occurred in July 2024, when the “blue screen of death” affected most Microsoft computers in airports and hospitals [24]. Although this incident was caused by a system bug in Microsoft software, its impact was similar to that of a denial-of-ML-service attack. Figure 5 shows MTG trees for three different impact techniques.

Refer to caption
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.

IV-C Numerical Experiments

This experiments section provides insight into the risk assessment process of penetration testing in an AI-enhanced hospital network. Assume the attacker chooses to erode the integrity of a regression ML model (using a support vector machine in this case) through a data poisoning algorithm as described in [11], which can degrade the model’s accuracy. We maintain the hospital network configuration as shown in Figure 3 and assume this ML model is placed in the AI center node of subnet B as the critical asset of the network. Let the risk score ω(vξ)𝜔conditional𝑣𝜉\omega(v\mid\xi)italic_ω ( italic_v ∣ italic_ξ ) of the entry node be directly identical to the data poisoning ratio, and denote the network uses non-sophisticated defenders that only use fixed strategies. By iterating the attacker’s skill level from 00 to 1111 with an interval of 0.10.10.10.1, we obtain the experimental results shown in the left of Figure 6.

Refer to caption
Figure 6: Left: Experimental results for the network risk score and ML model accuracy under different attacker’s skill levels with a fixed strategy defender. The blue line represents the network risk score, while the red line indicates the ML model accuracy. Right: The optimal attacker’s value of the entry node under different attacker’s skill levels. The orange line represents the fixed strategy defender, while the green line represents the purple teaming defender.

We show all possible risk scores and model accuracy at different levels of the attacker’s skill. As the attacker’s skill level increases, the risk score rises, leading to a corresponding decrease in the ML model’s accuracy. For a ca=1subscript𝑐𝑎1c_{a}=1italic_c start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT = 1 attacker, they can obtain nearly 100%percent100100\%100 % of the value from the critical asset and cause complete damage to the ML model.

The experiment provides a risk evaluation that helps users determine whether they need to increase their investment in preparation for future attacks and illustrates the impact of reachability on hospital security. For instance, refer to the left of Figure 6, if a user can tolerate a maximum accuracy loss of around 20%percent2020\%20 % for the ML model, they will likely increase their defense budget if they believe that future attackers will have a high skill level, specifically when ca0.4subscript𝑐𝑎0.4c_{a}\geq 0.4italic_c start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ≥ 0.4, vice versa. This demonstrates the vulnerability of a fixed-strategy defender when confronted with a strategic attacker. Let us consider a more strategic defender, such as a purple teaming defense, which is a collaborative cybersecurity strategy that integrates both offensive (red teaming) and defensive (blue teaming) perspectives to strengthen the overall security posture of the system. Suppose we aim to use a Stackelberg equilibrium to find an optimal purple teaming meta-penetration playbook. Let σav,subscriptsuperscript𝜎𝑣𝑎\sigma^{v,*}_{a}italic_σ start_POSTSUPERSCRIPT italic_v , ∗ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT be the anticipated optimal penetration plan for the attacker given the defense plan (it can be a fixed defense strategy or the defender’s current strategy). The optimal purple teaming defense plan σdv,pursubscriptsuperscript𝜎𝑣𝑝𝑢𝑟𝑑\sigma^{v,pur}_{d}italic_σ start_POSTSUPERSCRIPT italic_v , italic_p italic_u italic_r end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT is given by σdv,pur(σcv)maxσdvΣdvudv(σav,,σdv,σcv)subscriptsuperscript𝜎𝑣𝑝𝑢𝑟𝑑subscriptsuperscript𝜎𝑣𝑐subscriptsubscriptsuperscript𝜎𝑣𝑑subscriptsuperscriptΣ𝑣𝑑subscriptsuperscript𝑢𝑣𝑑subscriptsuperscript𝜎𝑣𝑎subscriptsuperscript𝜎𝑣𝑑subscriptsuperscript𝜎𝑣𝑐\sigma^{v,pur}_{d}(\sigma^{v}_{c})\in\max_{\sigma^{v}_{d}\in\Sigma^{v}_{d}}% \quad u^{v}_{d}(\sigma^{v,*}_{a},\sigma^{v}_{d},\sigma^{v}_{c})italic_σ start_POSTSUPERSCRIPT italic_v , italic_p italic_u italic_r end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT ( italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) ∈ roman_max start_POSTSUBSCRIPT italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT ∈ roman_Σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT end_POSTSUBSCRIPT italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT ( italic_σ start_POSTSUPERSCRIPT italic_v , ∗ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) subject to σav,argmaxσavΣavuav(σav,σdv,σcv)subscriptsuperscript𝜎𝑣𝑎subscriptsubscriptsuperscript𝜎𝑣𝑎subscriptsuperscriptΣ𝑣𝑎subscriptsuperscript𝑢𝑣𝑎subscriptsuperscript𝜎𝑣𝑎subscriptsuperscript𝜎𝑣𝑑subscriptsuperscript𝜎𝑣𝑐\sigma^{v,*}_{a}\in\arg\max_{\sigma^{v}_{a}\in\Sigma^{v}_{a}}u^{v}_{a}(\sigma^% {v}_{a},\sigma^{v}_{d},\sigma^{v}_{c})italic_σ start_POSTSUPERSCRIPT italic_v , ∗ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ∈ roman_arg roman_max start_POSTSUBSCRIPT italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ∈ roman_Σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT end_POSTSUBSCRIPT italic_u start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT ( italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_a end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT , italic_σ start_POSTSUPERSCRIPT italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ).

It is evident that the purple teaming defense is more sophisticated than a fixed strategy. The experimental results shown in the right of Figure 6 demonstrate the effectiveness of the purple teaming defense with Vmax=100subscript𝑉𝑚𝑎𝑥100V_{max}=100italic_V start_POSTSUBSCRIPT italic_m italic_a italic_x end_POSTSUBSCRIPT = 100 and a penalty of 5050-50- 50 for staying in a node. Regardless of the attacker’s skill level, their values always remain below 00, indicating that the defender has the advantage. As a trade-off, the purple teaming defender requires more computing resources. In the hospital context, a purple teaming defender can be seen as an upgrade from a fixed strategy defender, meaning the hospital has invested more budget in cybersecurity defense.

V Conclusion

In this work, we propose ADAPT, a game-theoretic and neuro-symbolic framework for automated distributed adaptive penetration testing. ADAPT addresses the critical demands of modern healthcare infrastructure and has the ability to dynamically adjust to evolving threats, and its effectiveness in securing AI systems against zero-day vulnerabilities underscores its potential as a vital tool for enhancing cybersecurity in healthcare environments. The case study demonstrates a real-life hospital network structure and showcases the framework’s effectiveness in addressing various tactical techniques employed by attackers. In numerical experiments, the results show around 98% improvement in reducing the risk score of the system by using the purple teaming defender as the replacement of a fixed strategy defender.

References

  • [1] M. Samorani, S. Harris, L. G. Blount, H. Lu, and M. A. Santoro, “Overbooked and overlooked: Machine learning and racial bias in medical appointment scheduling,” Manufacturing & Service Operations Management, p. 19, 2021. [Online]. Available: https://doi.org/10.2139/ssrn.3471420
  • [2] M. Eshghali, D. Kannan, N. Salmanzadeh-Meydani, and A. M. E. Sikaroudi, “Machine learning based integrated scheduling and rescheduling for elective and emergency patients in the operating theatre,” Annals of Operations Research, vol. 332, no. 1, pp. 989–1012, 2024. [Online]. Available: https://doi.org/10.1007/s10479-023-05168-x
  • [3] M. M. Ahsan, S. A. Luna, and Z. Siddique, “Machine-learning-based disease diagnosis: A comprehensive review,” Healthcare (Basel), vol. 10, no. 3, p. 541, 2022. [Online]. Available: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8950225/
  • [4] T. J. Loftus, A. C. Filiberto, Y. Li, and et al., “Decision analysis and reinforcement learning in surgical decision-making,” Surgery, vol. 168, no. 2, pp. 253–266, 2020.
  • [5] S. Gorbunov and A. Rosenbloom, “Autofuzz: Automated network protocol fuzzing framework,” 2010. [Online]. Available: https://api.semanticscholar.org/CorpusID:18430752
  • [6] Y. Stefinko, A. Piskozub, and R. Banakh, “Manual and automated penetration testing: Benefits and drawbacks,” in 2016 13th International Conference on Modern Problems of Radio Engineering, Telecommunications and Computer Science (TCSET).   IEEE, 2016, pp. 488–491.
  • [7] M. C. Ghanem and T. M. Chen, “Reinforcement learning for efficient network penetration testing,” Information, vol. 11, no. 1, p. 6, 2019.
  • [8] Z. Hu, R. Beuran, and Y. Tan, “Automated penetration testing using deep reinforcement learning,” in 2020 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW), 2020, pp. 2–10.
  • [9] D. Shmaryahu, G. Shani, J. Hoffmann, and M. Steinmetz, “Partially observable contingent planning for penetration testing,” in IWAIS: First International Workshop on Artificial Intelligence in Security, 2017.
  • [10] B. Mueller, “Understanding and mitigating the risk of ai vs. traditional software,” 2023. [Online]. Available: https://www.isaca.org/resources/news-and-trends/industry-news/2023/understanding-and-mitigating-the-risk-of-ai-vs-traditional-software
  • [11] F. A. Yerlikaya and S. Bahtiyar, “Data poisoning attacks against machine learning algorithms,” Expert Systems with Applications, vol. 208, p. 118101, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0957417422012933
  • [12] R. R. Wiyatno, A. Xu, O. Dia, and A. de Berker, “Adversarial examples in modern machine learning: A review,” 2019. [Online]. Available: https://arxiv.org/abs/1911.05268
  • [13] T. Li, H. Lei, and Q. Zhu, “Sampling attacks on meta reinforcement learning: A minimax formulation and complexity analysis,” 2023. [Online]. Available: https://arxiv.org/abs/2208.00081
  • [14] L. Head, “Rising threat of ddos attacks in healthcare,” 2024. [Online]. Available: https://aisn.net/rising-threat-of-ddos-attacks-in-healthcare/
  • [15] Skylight Cyber, “Cylance, i kill you!” 2019. [Online]. Available: https://skylightcyber.com/2019/07/18/cylance-i-kill-you/
  • [16] E. M. Hutchins, M. J. Cloppert, and R. M. Amin, “Intelligence-driven computer network defense informed by analysis of adversary campaigns and intrusion kill chains,” in Proceedings of the 2010 IEEE Symposium on Security and Privacy, 2010. [Online]. Available: https://api.semanticscholar.org/CorpusID:6421896
  • [17] A. E. LeBouthillier, “Symbolic artificial intelligence and first order logic,” 1999. [Online]. Available: https://home.csulb.edu/ wmartinz/content/symbolic-artificial-intelligence-and-first-order-logic.html
  • [18] A. Salleh, “Network architecture for healthcare information systems,” 2014. [Online]. Available: https://drdollah.com/hospital-information-system-his/system-architecture/
  • [19] O. H. USA, “Building hardware architecture for healthcare network computing for internet of things and artificial intelligence,” 2024. [Online]. Available: https://www.onyxhealthcareusa.com/building-hardware-architecture-for-healthcare-network-computing-for-internet-of-things-and-artificial-intelligence/
  • [20] P. Krass, “A hospital’s diagnosis: Professional ai workloads require professional hardware,” 2023. [Online]. Available: https://www.performance-intensive-computing.com/objectives/a-hospital-s-diagnosis-professional-ai-workloads-require-professional-hardware/
  • [21] MITRE, “Mitigations enterprise mitre att&ck,” Bedford, MA, USA, 2020. [Online]. Available: https://attack.mitre.org/mitigations/enterprise/
  • [22] MITRE, “Atlas (adversarial threat landscape for artificial-intelligence systems),” Bedford, MA, USA, 2024. [Online]. Available: https://atlas.mitre.org/
  • [23] R. Lakshmanan, “New attack technique ’sleepy pickle’ targets machine learning models,” 2024. [Online]. Available: https://thehackernews.com/2024/06/new-attack-technique-sleepy-pickle.html
  • [24] The New York Times, “Chaos and confusion: Tech outage causes disruptions worldwide,” 2024. [Online]. Available: https://www.nytimes.com/2024/07/19/business/microsoft-outage-cause-azure-crowdstrike.html