Anti-central Nodes in Networks

• Anti-central nodes are network elements that may display low traditional centrality while playing crucial roles in hidden or context-dependent applications, such as AML detection and adversarial networks.
• They are employed to optimize coverage and robustness in decentralized federated learning and ISP vulnerability studies, challenging simple hub-based approaches.
• Empirical studies in control, spreading dynamics, and FBAS demonstrate that anti-central nodes can either act as stealth influencers or low-impact peripherals, guiding diverse strategic interventions.

Anti-central nodes are not a single universally defined network object. Across recent work, the term denotes several related but distinct phenomena: nodes that are structurally important while remaining inconspicuous under standard centrality measures; nodes deliberately placed away from global hubs to maximize coverage or reduce redundancy; nodes that are simply least central under a chosen metric and whose removal causes minimal disruption; and nodes that are globally peripheral yet dynamically important at particular spatial or temporal scales. The common thread is a departure from naive “highest-centrality-node” reasoning, either because importance is hidden, distributed, context-dependent, or explicitly defined as the complement of centrality (Bonato et al., 12 Sep 2025, Piaseczny et al., 10 Nov 2025, Bonato et al., 2022, Gaul et al., 2020).

1. Conceptual scope and principal meanings

The literature uses “anti-central” in at least five technically distinct ways. In anti-money-laundering detection, anti-central nodes are accounts that play a structurally important role while avoiding high degree, high betweenness, and similar classical signatures (Bonato et al., 12 Sep 2025). In adversarial networks, the closely related construct is the low-key leader: a node with high CON score but low PageRank on the reversed graph, hence influential without appearing visibly central (Bonato et al., 2022). In decentralized federated learning, anti-centrality appears as a placement principle: adversaries should not be piled onto the same globally central hub region, but spread across the graph to avoid overlap in influence (Piaseczny et al., 10 Nov 2025). In ISP vulnerability analysis, the paper does not coin the term, but the natural contextual interpretation is the bottom of the centrality ranking: peripheral nodes whose removal produces minimal changes in connectivity and traffic metrics (Nomikos et al., 2013). In FBAS analysis, low intactness-based centrality identifies nodes whose becoming ill-behaved rarely causes others to become befouled, which is the most direct robustness-oriented notion of anti-centrality in that model (Gaul et al., 2020).

Context	Operational meaning of anti-centrality	Representative paper
AML transaction cycles	Hidden importance with low classical centrality	(Bonato et al., 12 Sep 2025)
Adversarial networks	High CON, low PageRank “low-key leaders”	(Bonato et al., 2022)
Decentralized FL	Coverage-first, anti-cluster placement away from hub concentration	(Piaseczny et al., 10 Nov 2025)
ISP vulnerability	Least-central, low-impact peripheral nodes	(Nomikos et al., 2013)
FBAS	Low intactness-based ability to befoul others	(Gaul et al., 2020)

This plurality matters because anti-centrality is sometimes an indicator of hidden importance and sometimes an indicator of low importance. Confusing these two usages is a recurring source of misunderstanding.

2. Hidden importance: anti-centrality beyond classical graph metrics

In anti-money-laundering detection, anti-central nodes are explicitly defined as accounts that are structurally important yet “do not look central” under degree, betweenness, closeness, and related measures. The setting is an anonymized Rabobank transaction network modeled as a directed graph with $1{,}048{,}166$ nodes and $1{,}686{,}184$ directed edges after filtering; the analysis focuses on 83 previously detected directed cycles involving 155 accounts (Bonato et al., 12 Sep 2025). The paper operationalizes anti-centrality at the cycle level through a node2vec-based spread number $r(C)$ and an aggregate score

$$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$

after min–max normalization of the components. High $R(C)$ corresponds to cycles that are embedding-dispersed yet low in community-normalized betweenness and degree. Only 19 of the 83 cycles have non-zero spread number, and the top three cycles by $R$ involve only 7 accounts because of overlap. No supervised performance metrics such as precision, recall, or AUC are reported, since no ground-truth labels are available; the evaluation is structural and interpretive rather than predictive (Bonato et al., 12 Sep 2025).

A closely related formulation appears in adversarial networks. There, anti-central nodes correspond to low-key leaders: nodes with high CON score and low PageRank on the reversed-edge network. After unity-based normalization, low-key leader strength is defined as

$$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$

and a node is considered a low-key leader if it attains the maximum $\varepsilon_i$ and $\varepsilon_i>0.5$ (Bonato et al., 2022). Empirically, the paper reports that 155 of 172 animal dominance networks contain at least one such node under the $0.5$ threshold, and 95.35% do under a $1{,}686{,}184$0 threshold. The same contrast appears in G20 trade data and Bitcoin mistrust networks. In both the AML and low-key-leader settings, anti-centrality is not low importance; it is concealed importance.

This suggests a general pattern: when actors have incentives to avoid detection, classical centrality may systematically miss the nodes of greatest structural interest. In such cases, anti-centrality is better understood as a discrepancy between visible centrality and latent structural role.

3. Coverage, dispersion, and anti-cluster placement

In decentralized federated learning, anti-centrality appears not as a node attribute alone but as a set-level placement strategy. The system is modeled as a strongly connected, time-invariant directed communication graph $1{,}686{,}184$1, with honest nodes and adversarial nodes partitioning $1{,}686{,}184$2, and the learning algorithm is S-AB gradient tracking. The paper studies coordinated adversarial node placement under local FGSM-style poisoning and evaluates random placement, eigenvector-centrality-based placement, MaxSpAN-FL, and Hopping-Augmented MaxSpAN-FL (Piaseczny et al., 10 Nov 2025).

The conceptual pivot is that choosing the top-$1{,}686{,}184$3 nodes by eigenvector centrality can be redundant because central nodes often cluster in the same region of the graph. MaxSpAN-FL instead seeks to maximize total pairwise adversarial distance,

$1{,}686{,}184$4

using a BFS-cluster heuristic that minimizes overlap between adversarial influence regions. The resulting attack is “span-only”: it disperses adversaries irrespective of whether the selected nodes are globally most central. The hybrid Hopping-Augmented MaxSpAN-FL starts from that dispersed placement and then allows local stochastic hopping toward the most central neighbor, with hop probability controlled by degree variance, local clustering, decision parameters $1{,}686{,}184$5, and decay $1{,}686{,}184$6. The paper’s interpretation is explicit: MaxSpAN-FL is “anti-cluster,” and in heterogeneous graphs this often means selecting nodes that are not the very highest in global centrality but are separated local centers (Piaseczny et al., 10 Nov 2025).

The empirical results are topology dependent. On directed geometric graphs, MaxSpAN-FL and Hopping-Augmented MaxSpAN-FL outperform eigenvector-centrality-based attacks, and eigenvector centrality often performs worse than random placement. On Erdős–Rényi graphs, all attacks perform similarly. On sparse preferential-attachment graphs with few adversaries, pure eigenvector-centrality placement can be strongest, but as density or adversarial fraction increases, dispersion becomes more important; in very dense PA graphs, Hopping-Augmented and MaxSpAN-FL outperform random by about $1{,}686{,}184$7–$1{,}686{,}184$8, and in larger DG graphs they often achieve AAL advantages greater than $1{,}686{,}184$9 over random. One reported 25-node, 2-adversary case shows an improvement over the next best attack of about $r(C)$0 (Piaseczny et al., 10 Nov 2025).

The theoretical analysis supports the empirical claim that pure eigenvector centrality is suboptimal. Under the simplified consensus model $r(C)$1, the lower bound on attack impact contains both adversarial and honest-node terms weighted by eigenvector centralities. Increasing the adversarial $r(C)$2 does not necessarily maximize the bound because the honest-node contribution changes as well. The paper’s conclusion is therefore set-level rather than node-level: effective adversarial placement is a coverage problem, not a simple top-centrality ranking problem (Piaseczny et al., 10 Nov 2025).

4. Peripherality, dispensability, and low-impact nodes

A different meaning of anti-centrality is the literal opposite of centrality. In the ISP vulnerability study, the paper does not use the term “anti-central,” but its methodology makes the interpretation straightforward: nodes at the bottom of centrality rankings are those whose removal causes minimal degradation of the size of the giant connected component, the number of connected components, average shortest path length, and aggregate maximum flow (Nomikos et al., 2013). The paper studies seven indices—degree, betweenness, closeness, harmonic, eccentricity, eigenvector, and PageRank—and shows that top-$r(C)$3 overlap predicts attack impact better than full-ranking correlation. It also shows that eccentricity-based removal is consistently least harmful, while degree often approximates the most damaging strategies for traffic-carrying capacity. In this framework, anti-central nodes are leaves, near-leaf routers, and peripheral nodes that are low in several centrality rankings and empirically safe to fail (Nomikos et al., 2013).

In FBAS analysis, the most problem-adapted notion of anti-centrality is based on intactness. An FBAS is a pair $r(C)$4 with quorum slices $r(C)$5, quorums $r(C)$6, and the key safety property of quorum intersection. A set $r(C)$7 is a DSet if the reduced system $r(C)$8 has quorum intersection and either $r(C)$9 is a quorum in $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$0 or $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$1. Given a set $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$2 of ill-behaved nodes, a node $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$3 is $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$4-intact if there exists a DSet $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$5 with $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$6; otherwise it is $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$7-befouled (Gaul et al., 2020).

This leads to intactness-based centralities. The linear-system variant uses

$$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$8

where $$R(C) = \frac{1}{3}\left(r(C) + \big(1-\mathrm{CNS}(C,B)\big) + \big(1-\mathrm{CNS}(C,D)\big)\right),$$9 aggregates, over weighted ill-behaved sets $R(C)$0 containing $R(C)$1, whether $R(C)$2 becomes befouled. The hierarchical variant uses a matrix $R(C)$3 that restricts attention using reachability in the trust graph, yielding

$R(C)$4

After normalization, low $R(C)$5 or $R(C)$6 means that a node’s becoming ill-behaved rarely causes other nodes to lose intactness. That is the most direct FBAS analogue of an anti-central node (Gaul et al., 2020).

The examples make the distinction concrete. In one 5-node example, $R(C)$7 ranks node 2 lowest, whereas trust-graph and quorum-based measures miss or flatten that distinction. In the 7-node example with unique quorum-intersection node 7, trust-graph subgraph centrality assigns the lowest value to node 7 even though node 7 is maximally safety-critical; intactness-based and quorum-based centralities correctly place it highest. In Stellar-like examples, nodes in organization A are highest and nodes in C lowest under $R(C)$8 and especially $R(C)$9, matching the intended notion of structural dispensability. A notable counterexample is the “new-2” Stellar-like network, where the only DSets are $R$0 and $R$1; then $R$2 for all nodes, so no node is anti-central in the intactness sense because every node is essential to safe operation (Gaul et al., 2020).

5. Dynamical, control-theoretic, and spreading interpretations

Anti-centrality also appears in dynamical settings where global graph centrality is not the right criterion for influence or controllability. In the study of time-varying control schedules, the relevant quantity is the scale-dependent $R$3-communicability

$R$4

For unweighted networks, $R$5 is degree-like, while $R$6 ranks nodes by left eigenvector centrality. If the maximizers of $R$7 and $R$8 differ, then for sufficiently large horizon $R$9 the network is in class V and time-varying control outperforms time-invariant control. The optimal policy is

$$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$0

so globally central nodes are actuated early and locally central nodes late (Nozari et al., 2016). In this framework, some nodes are anti-central globally yet central at short scales. The paper reports $$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$1 for the air500 network and $$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$2 for the WesternUS power grid under the trace metric. This suggests that globally peripheral nodes can be operationally central near the control horizon.

A related control-theoretic inversion appears in anti-stable networks. There, the best driver node is not selected by degree or betweenness but by maximizing the volume of the region of attraction under bounded control. For the fully anti-stable case,

$$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$3

and the preferred driver is $$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$4 (Mahia et al., 2014). The paper does not use the term anti-central explicitly, but it emphasizes that the best driver node is central in the unstable eigenspace rather than in static graph topology. In a 2-node example, two structurally symmetric nodes yield different attraction areas; in a 4-node mixed-spectrum example, node 4 gives the largest area even though structural connectivity alone does not determine this choice (Mahia et al., 2014).

Spreading processes produce yet another operational notion. In the anti-rumor model, the main design questions are timing and seed selection. The reported simulations show a timing threshold $$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$5: on the Enron email network, $$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$6 time steps for both APR and APR-PHB, while in BA scale-free networks the timing threshold decreases as average degree increases until close to zero. For seed selection, coreness is a better topological descriptor than degree for identifying influential anti-rumor nodes, and higher coreness corresponds to smaller final rumor size when used as an anti-rumor seed (Ji et al., 2013). This is not “anti-centrality” in the low-centrality sense; rather, it identifies nodes central for counter-spreading. A plausible implication is that the term “anti-central” becomes especially unstable in dynamical settings unless one specifies the underlying task—control, attack, or mitigation.

6. Methods, limitations, and recurring misconceptions

Several methodological points recur across the literature. First, anti-centrality is often task specific. In AML and adversarial networks, it is defined by a mismatch between latent structural importance and standard centrality (Bonato et al., 12 Sep 2025, Bonato et al., 2022). In decentralized FL, it is a dispersion principle over sets of nodes rather than an intrinsic scalar attached to an individual node (Piaseczny et al., 10 Nov 2025). In ISP vulnerability and FBAS, it can mean low-impact or structurally dispensable nodes (Nomikos et al., 2013, Gaul et al., 2020). In control, a globally modest node may still be optimal at a specific scale or under saturation constraints (Nozari et al., 2016, Mahia et al., 2014).

Second, cheap graph proxies can be misleading. Trust-graph subgraph centrality in FBAS can assign the lowest value to the unique quorum-intersection node, because it measures closed walks rather than safety dependence (Gaul et al., 2020). Eigenvector-centrality placement in decentralized FL can underperform random placement on geometric graphs because it clusters attackers into overlapping regions (Piaseczny et al., 10 Nov 2025). Full-ranking correlation between centrality indices in ISP networks can be high while top-$$\varepsilon_i = \mathrm{CON}_{i,\mathrm{norm}} - \mathrm{PR}_{i,\mathrm{norm}},$$7 overlap is modest, and the latter better predicts vulnerability (Nomikos et al., 2013).

Third, the computational burden is highly uneven. AML anti-centrality relies on repeated node2vec runs and community-normalized scores, but no ground-truth labels are available for direct supervised validation (Bonato et al., 12 Sep 2025). FBAS intactness-based measures depend on computing DSets and quorum intersection, and the decision problem for quorum intersection is NP-complete (Gaul et al., 2020). Global graph-based centralities in ISP and control settings can also be expensive, whereas local degree can sometimes approximate worst-case traffic vulnerability surprisingly well (Nomikos et al., 2013).

A final misconception is to treat anti-central nodes as necessarily benign or necessarily peripheral. The literature does not support that simplification. Some anti-central nodes are stealth influencers, some are low-impact leaves, some are locally central but globally peripheral actuators, and some are deliberately selected non-hubs whose collective placement dominates the system more effectively than attacking hubs would. The most defensible general statement is therefore narrow: anti-central nodes are nodes whose relevance is not captured by straightforward global-centrality ranking, either because their importance is hidden, context dependent, or explicitly defined as low structural criticality.