Localized Attack Vulnerability Index

• Localized Attack Vulnerability Index (LAVI) is a metric that quantifies network vulnerability under targeted, spatially localized node removals by measuring cumulative link fragmentation.
• It employs both BFS-based, shell-ordered removal sequences and percolation theory approaches to identify critical nodes and assess global connectivity impacts.
• Empirical results demonstrate that LAVI outperforms classical centrality metrics in predicting network robustness in diverse synthetic and real-world network models.

The Localized Attack Vulnerability Index (LAVI) is a quantitative metric for assessing the susceptibility of complex networks to spatially-constrained node removal processes, known as localized attacks (LA). Unlike classical centrality metrics designed for random or global perturbations, LAVI captures how the structural position of an initial failure and the dynamic propagation of subsequent removals jointly determine the global impact on connectivity. Two distinct but complementary mathematical formulations exist: a node-level, seed-specific LAVI (for assessing “worst-case” or critical locations) (Chujyo et al., 6 Feb 2026), and a percolation-theoretic, network-level LAVI grounded in ensemble-averaged random graph models (&&&1&&&).

1. Formal Definitions

Node-level LAVI (Seed-based Localized Attack)

Let $G = (V, E)$ be an undirected network. For each seed node $i \in V$, the LAVI $\mathcal{L}(i)$ quantifies the cumulative number of half-links severed as a localized attack propagates shell-by-shell outward from $i$:

• For each node $v \in V$, compute the shortest-path distance $d_i(v) = \mathrm{dist}_G(i, v)$.
• Define shells: $S_\ell = \{ v : d_i(v) = \ell \}$, $\ell = 0, 1, ..., d_{\mathrm{max}}$.
• Construct a removal sequence $\pi(i)$ by concatenating shells in increasing distance order. Within each shell, nodes may be randomly shuffled (“random-case”) or sorted by descending original degree (“worst-case”).
• As each node $v_k$ is removed (in order), update the cumulative severed half-link count $C_k$: $C_0 = 0$, $C_k = C_{k-1} + \deg_G(v_k)$.
• LAVI is the normalized area under the $C_k$ curve:

$$\mathcal{L}(i) = \frac{1}{|V|} \sum_{k=1}^{|V|} C_k = \frac{1}{|V|} \sum_{k=1}^{|V|} (|V| - k + 1) \deg_G(v_k)$$

A larger $\mathcal{L}(i)$ indicates more severe and rapid link fragmentation, reflecting higher vulnerability associated with the seed node $i$.

Network-level LAVI (Percolation-based Formulation)

Given a random graph model with degree distribution $P(k)$ and generating function $G_0(x)$:

• Perform a localized attack removing a connected fraction $1-p$ of nodes.
• After “Stage I–II” formalism (Shao et al., 2014), new distributions are defined by $P_p(k) = P(k) f^k / G_0(f)$, with $p = G_0(f)$ and $\tilde{p} = G_0'(f) / (f G_0'(1))$.
• The critical fraction $p_c^{\mathrm{loc}}$ for global connectivity is solved by $G_0''(f_c) = G_0'(1)$, with $p_c^{\mathrm{loc}} = G_0(f_c)$.
• Mean vulnerability just above threshold is $\overline{V} = 1 - G_0(u(p))$, where $u(p)$ is the non-giant stub probability solution: $u = G_1^p(u) = [G_0^p]'(u)/[G_0^p]'(1)$.
• LAVI is defined as

$$\mathrm{LAVI} = \frac{\overline{V}(p^*)}{p_c^{\mathrm{loc}}}$$

where $p^* > p_c^{\mathrm{loc}}$ is a small margin above the threshold.

2. Computational Procedures

Node-level LAVI Computation

1. Breadth-first search (BFS) from seed $i$ to obtain $d_i(v)$ for all $v$.
2. Shell grouping: Partition $V$ into $\{S_0, S_1, ..., S_{d_{\max}}\}$.
3. Within-shell ordering: Random shuffle for “random-case”; descending degree for “worst-case”.
4. Sequence aggregation: Concatenate ordered shells to $[v_1, ..., v_{|V|}]$.
5. Cumulative calculation: For $k=1$ to $|V|$, accumulate $C_k$ and total sum.
6. Normalization: Return $\mathcal{L}(i) = \mathrm{total}/|V|$.

Algorithmic complexity is $O(|E| + |V| \log |V|)$ per node.

Percolation-based LAVI Computation

1. Solve $G_0''(f_c) = G_0'(1)$ for $f_c$.
2. Compute $p_c^{\mathrm{loc}} = G_0(f_c)$.
3. Form $G_0^p(x)$ at chosen $p > p_c^{\mathrm{loc}}$.
4. Solve self-consistent $u = G_1^p(u)$.
5. Evaluate $\overline{V} = 1 - G_0(u)$.
6. Output $\mathrm{LAVI} = \overline{V}/p_c^{\mathrm{loc}}$.

These steps use only the degree distribution and generating function, facilitating analysis for arbitrary random graph ensembles.

3. Theoretical Motivation and Distinctions

Classical global centrality measures (degree, closeness, betweenness) assess node importance under unstructured failure, but localized attacks spread shell by shell from the seed, making the sequence and location of removals essential to global damage. LAVI explicitly integrates the cumulative severing of links over the removal process, weighting early-stage removals—where fragmentation impacts network robustness nonlinearly—more heavily. This design aligns with percolation theory insights that early-loss links disproportionately influence the onset of connectivity breakdown (Chujyo et al., 6 Feb 2026, Shao et al., 2014).

The percolation-theoretic LAVI generalizes this insight to network ensembles, identifying two principal signals of vulnerability: (i) a low $p_c^{\mathrm{loc}}$ (collapse at low remaining fraction), and (ii) high mean vulnerability $\overline{V}$ near threshold (substantial population excluded from the giant).

4. Empirical Evidence and Numerical Results

Synthetic Networks

Barabási–Albert (BA) networks with $m=2$ are used at sizes $N = 10^2, 10^3, 10^4$. For each $N$, seed-averaged robustness index $R$ (mean largest connected component fraction) is estimated after 100 LA trials per seed.

Key results (Spearman $\rho$/Pearson $r$):

Metric	$N=100$: $\rho$/$r$	$N=1000$: $\rho$/$r$	$N=10000$: $\rho$/$r$
$\mathcal{L}_{\mathrm{random}}$	–0.913/–0.900	–0.693/–0.717	–0.570/–0.590
$\mathcal{L}_{\mathrm{worst}}$	–0.818/–0.788	–0.663/–0.692	–0.464/–0.488
degree	0.140/0.043	–0.031/–0.033	–0.033/0.013
closeness	0.043/0.082	–0.009/–0.083	0.181/0.138
betweenness	0.037/0.036	–0.081/–0.062	0.062/0.023

$\mathcal{L}_{\mathrm{random}}$ consistently outperforms classical centralities in predicting global robustness degradation under LA (Chujyo et al., 6 Feb 2026).

Real-world Networks

Seven diverse networks: chemistry, power grid, biology, infrastructure, and social. For instance,

• power-494-bus network: Spearman $\mathcal{L}_{\mathrm{random}} = -0.81$, $\mathcal{L}_{\mathrm{worst}} = -0.85$, versus degree $-0.11$.
• inf-openflights network: Pearson $\mathcal{L}_{\mathrm{random}} = -0.97$, $\mathcal{L}_{\mathrm{worst}} = -0.98$.

LAVI provides the strongest (most negative) correlation with $R$ in 6 out of 7 networks. Exceptional cases, such as the bio-diseasome network with only $\sim-0.10$ correlation, are due to fragmentation governed by meso-scale bridges outside the scope of localized shell propagation (Chujyo et al., 6 Feb 2026).

Percolation-based LAVI

Analytical and numerical analysis yields:

• Erdős–Rényi: $p_c^{\mathrm{loc}} = 1/\langle k \rangle$. LAVI scales as $\langle k \rangle^2 \delta$ for an overshoot $\delta$ above threshold.
• Random-Regular: $p_c^{\mathrm{loc}} \approx 1/(k_0-1)$.
• Scale-Free: For $\gamma \in (2, 3)$, $p_c^{\mathrm{loc}}$ is small, but $\overline{V}$ jumps rapidly, and LAVI peaks for intermediate $\gamma$ (Shao et al., 2014).

5. Comparative Analysis with Classical Metrics

LAVI reveals vulnerability not detected by degree, closeness, or betweenness. In empirical experiments, scatter plots of seed robustness $R$ versus LAVI metrics show tight negative correlation (linear alignment) for LAVI but not for centrality. In topologies with high degree heterogeneity, early removal of high-degree nodes by LA sharply increases vulnerability, an effect captured by LAVI’s area-under-curve construction but missed by static centrality.

Classical centrality metrics assume attacks are global or independent, overlooking dynamic propagation features fundamental to localized threats (Chujyo et al., 6 Feb 2026).

6. Implications for Network Design and Resilience

High-LAVI nodes mark regions whose localized failure is most detrimental to structural robustness. Applications include:

• Critical infrastructure: Targeted addition of redundancy or reinforcement to high-LAVI zones in power grids, transportation, or communication networks.
• Resilience assessment: LAVI can quantify priorities for intervention against spatial threats (e.g., earthquakes, floods), outperforming traditional centralities.
• Adversarial contexts: High-LAVI nodes are optimal dismantling seeds for efficient network fragmentation.

Incorporating LAVI into analysis enables strategies that are specifically designed for propagation-aware mitigation, filling a methodological gap between static vulnerability assessment and the intrinsically dynamic nature of spatially localized threats (Chujyo et al., 6 Feb 2026).

7. Connections to Percolation Theory and Broader Context

LAVI is rooted in network percolation theory (Shao et al., 2014), explicitly modeling the abruptness of connectivity loss as cascades propagate from a seed. Percolation-based LAVI offers ensemble-level (graph type-dependent) benchmarks, while node-level LAVI provides actionable, fine-grained vulnerability maps for arbitrary finite networks. A plausible implication is that combining these approaches can give both “global” (ensemble) and “local” (seed-specific) perspectives on network resilience to spatially-structured failure, a scenario ubiquitous in real-world threats such as cascading outages and geographically localized hazards.