Attack Complexity in Security Systems

• Attack Complexity is a measure of the resources—computational steps, data queries, and iterations—required by an adversary to compromise security.
• It is applied across domains including cryptanalysis, machine learning robustness, and cyber-physical systems using frameworks like attack trees and complexity-theoretic models.
• Empirical metrics, such as CVSS Access Complexity and algorithmic runtime measures, offer practical insights into the exploitability and defense needs of systems.

Attack complexity encompasses the computational, data, or resource expenditure required for an adversary to successfully compromise, subvert, or degrade the target in the presence of defenses, protocol constraints, or architectural barriers. The precise meaning and formalization of attack complexity is context-dependent: it may refer to worst-case, average-case, or minimal attacker cost quantified in computational steps, run-time, iterations, query count, or information-theoretic resources. Attack complexity is a central concept across security evaluation, cryptanalysis, machine learning robustness, algorithmic defense, and cyber-physical threat modeling.

1. Formalizations of Attack Complexity

In foundational literature, attack complexity is typically encapsulated by resource-bounded measures tailored to the attack scenario:

• Combinatorial and Decision Theoretic: The minimal number of operations, samples, or iterations needed by an attacker to achieve a target misclassification, data reconstruction, or key recovery to within a specified tolerance or error threshold (Zhang et al., 16 Aug 2025, Pendyala, 2022). In machine learning or federated learning, for example, attack complexity may be formalized as the minimal number of attack iterations $T$ required to reduce the expected error below a given threshold $\tau$:

$$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$

(Zhang et al., 16 Aug 2025)

• Complexity-theoretic and Algorithmic: The number of logical, arithmetic, or oracle steps required under standard computational models (Turing machine, RAM, quantum circuit), considering worst-case and average-case performance. Classical time and space complexities (e.g., $O(N^2)$, $O(2^k)$) may refer to the attack algorithm’s asymptotics, the NP-hardness or inapproximability of attack planning, or the efficiency of cryptanalytic techniques (Audinot et al., 2017, Lopuhaä-Zwakenberg et al., 2022, Ercal, 2016, 0811.0063, Xu et al., 2021).
• Empirical/Heuristic Estimators: Attack complexity can also be measured by standardized indicators, e.g., CVSS Access Complexity scores (Low, Medium, High), as proxies for attacker effort in exploiting vulnerabilities (Allodi et al., 2018).

2. Attack Complexity in Security Modeling and Analysis

Attack complexity is analyzed and compared across formal frameworks, with notable rigor in:

• Attack Trees: In formal system modeling, attack trees describe the logical structure of multi-step attacks with AND/OR/SAND nodes. The complexity of verifying attack tree correctness (Meet, Match, Under-Match, Over-Match) is classified via complexity classes: for example, AND-nodes yield NP-complete or coNP-complete decision problems, whereas OR and SAND nodes are in PTIME (Audinot et al., 2017). Quantitative analysis of attack trees, such as computing the minimal-cost or $k$-top attacks, introduces further algorithmic complexity, which is linear in the tree size for static trees, and can be NP-hard for DAGs with shared substructures (Lopuhaä-Zwakenberg et al., 2022, Budde et al., 2021). Algorithms leveraging semiring domains or BDDs are used to manage these complexities.
• Network Robustness Metrics: Complexity-theoretic hardness is established for measures such as vertex attack tolerance (VAT) and unsmoothened VAT (UVAT), which quantify the minimum attacker effort needed to fragment a network beyond a given resilience threshold. Exact and approximate computation is NP-hard and even approximation within any constant or $n^{\delta}$ factor is impossible under standard complexity assumptions, delineating sharp barriers for attack-optimization strategies (Ercal, 2016).
• Graph-Based Attacks: In attacks on graph neural networks, computational complexity is dominated by the need to process large adjacency matrices. Traditional gradient-based attacks scale as $O(N^2 d^2)$ time and $O(N^2)$ space, but recent approaches (e.g., Simplified Gradient-based Attack, SGA) exploit locality to reduce attack complexity to $O(d^{3k} \Delta)$, where $\tau$0 is hop-radius, $\tau$1 is degree, and $\tau$2 is the number of edge modifications (Li et al., 2020).

3. Cryptanalytic Attack Complexity

Cryptanalytic frameworks provide classic taxonomies of attack complexity, quantifying effort in bit operations, time, memory, and data requirements:

• Key Recovery and Guess-and-Determine: For A5/1, the attack complexity of a new guess-and-determine strategy is calculated via an exact combinatorial analysis, yielding $\tau$3 operations and $\tau$4 GB memory—far less than $\tau$5 brute force—based on the average number of rounds required to reduce the candidate space to a small set for each guess of the R1 register. Trade-offs are formally analyzed in terms of key stream bits needed and resource optimization (Shah et al., 2012).
• RSA and Lattice-based Attacks: For variants of RSA attacks (Wiener/Dujella), attack complexity is expressed in terms of bounds on the secret exponent (e.g., $\tau$6), with classical methods requiring $\tau$7 time. Meet-in-the-middle strategies reduce the time complexity to $\tau$8 and space $\tau$9 via precomputation and hash table lookup, as opposed to quadratic brute-force search (0811.0063).
• Quantum Cryptanalysis: For quantum meet-in-the-middle and related-key attacks, complexity is redefined in terms of quantum oracle queries, superposition accesses, and amplitude-amplified search. For example, Bernstein–Vazirani-based related-key attacks can recover generic block-cipher keys in $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$0 quantum operations under quantum superposition access to the oracle, an exponential speedup over the $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$1 Grover bound (Xie et al., 2018). Quantum MITM on Feistel constructions yields $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$2 time for $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$3-round ciphers, by hybridizing inner Grover/amplitude amplification and outer Grover search (Xu et al., 2021). For lattice-based schemes and the LWE dual attack, quantum rejection sampling yields a quadratic speedup over classical MCMC, reducing the core attack term from $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$4 to $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$5, directly translating to security-level reductions for standard schemes such as Kyber (Ling et al., 24 May 2026).

4. Practical and Empirical Attack Complexity Metrics

Attack complexity in real systems is often tied to practical exploitability and vulnerability management:

• Vulnerability Exploitation: The “Attack Complexity” subscore in CVSS, along with “Impact,” drives empirical measures such as “Attack Potential,” which predicts the expected prevalence of attacks in the wild. Attack Complexity is crucial for distinguishing high-exploit potential vulnerabilities, with empirical data showing that low-complexity vulnerabilities are targeted orders of magnitude more frequently than high-complexity ones, regardless of impact (Allodi et al., 2018).
• Denial-of-Service (DoS) via Algorithmic Complexity: Attack complexity in CPU-based DoS is quantified by the attacker’s ability to force worst-case execution in algorithmic routines (e.g., modular exponentiation, hash-table collision, sorting) using minimal payload and bandwidth. Metrics such as Attacker’s Gain and Throttled Time are introduced to measure service degradation, while cost-benefit models inform scaling and mitigation strategies (Bhuiyan et al., 2022).
• SAT-based Hardware Attacks: In logic locking, attack complexity is typically measured in SAT-solver iterations to recover the secret key. For canonical XOR/XNOR-based lockings, each iteration halves the remaining key space, so attack complexity is $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$6 in key bits. For schemes employing point functions (e.g., AntiSAT, CAS-Lock), complexity can be tuned to $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$7 or $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$8 under certain key constraints, but generically remains linear unless structural redundancy is avoided (Zhong et al., 2022).

5. Attack Complexity in Control and Cyber-Physical Systems

Theoretical hardness results have been established for attack planning in estimation and control:

• Sensor Attack on Kalman Filtering: The problem of optimally removing sensors to maximize the steady-state covariance trace (i.e., degrade estimation accuracy) is shown to be not only NP-hard but also fundamentally inapproximable: no polynomial-time algorithm can guarantee even a constant-factor approximation unless P=NP. This stands in sharp contrast to other submodular sensor selection or actuation problems and is proved via reductions from the Exact Cover by 3-Sets (X3C) problem, with additional counterexamples showing that standard greedy heuristics perform arbitrarily poorly (Ye et al., 2020).
• Multi-Armed Bandit Manipulation: In adversarial attacks on UCB-based best-arm identification, sample complexity is precisely characterized: the round-complexity required for an adversary to promote a chosen arm $$S_k(\tau) = \min \left\{ T \mid \mathbb{E}\Big[\frac{1}{|D^{(k)}|} \sum_{i=1}^{|D^{(k)}|}\frac{1}{T}\sum_{t=1}^{T} d(\tilde X_{t,i}^{(k)}, X_i^{(k)}) \Big] \leq \tau \right\}.$$9 as best is $O(N^2)$0, scaling linearly in arms and sample noise, and quadratically in the inverse attack margin. The result delineates the minimal attack budget needed and shows that confidence-based learning is only as robust as the gap parameter and variance allow (Pendyala, 2022).

6. Synthesis: Factors Influencing Attack Complexity

Attack complexity fundamentally reflects the interplay between attacker algorithms, system structure, defenses, and information asymmetries:

• System model: Richer system semantics (state spaces, tree structures, network topologies) or design choices (shared substructures, nonlinearity) can significantly raise or lower attack complexity.
• Defensive mechanisms: Obfuscation, randomization, privacy-preserving noise, or gating mechanisms impact the minimal resources needed; e.g., MBP privacy increases attack complexity polynomially in the model dimension and inverse privacy budget (Zhang et al., 16 Aug 2025).
• Metric and objective: Whether the goal is key recovery, system degradation, data reconstruction, or service outage, the attack complexity metric (worst-case, average-case, adaptive) changes accordingly.
• Resource models: Classical, quantum, and stochastic resource models yield sharply different attack complexity regimes, with quantum access frequently offering dramatic (sometimes exponential) reductions in attack cost (Xie et al., 2018, Xu et al., 2021, Ling et al., 24 May 2026).
• Intractability results: Many natural attack optimization problems are not only NP-hard but non-approximable, implying a reliance on heuristics or parameterized hardness except in restricted domains (Ercal, 2016, Ye et al., 2020).

Attack complexity thus serves as a critical lens for evaluating system security, resilience, and vulnerability from both a theoretical and practical standpoint. It shapes formal verification, red-teaming, privacy engineering, cryptosystem selection, and deployment architecture decisions throughout academia and industry.