# A contribution to the critique of blockchain censorship

Crypto.world

> Earlier work (Yeo and Zhang, cited as [21]) showed that when validators arrive in random order and can coordinate through a single message on the blockchain, joining a censorship coalition is a…

**arXiv**: 2608.15640

**Published**: 2026-08-16

**Authors**: Ruichao Jiang, Michelle Yeo, Long Wen

## AI summary

A contribution to the critique of blockchain censorship

## Overview

**Research area:** Blockchain security and mechanism design — specifically consensus-level censorship attacks on proof-of-stake networks (cs.CR).

**Technical level:** Advanced. The paper is built on probability theory (hypergeometric order statistics, strong concavity, Jensen's inequality), formal game-theoretic equilibrium reasoning, and Monte Carlo simulation, and assumes familiarity with validator voting power and block-reward mechanics.

**Scope (one sentence):** The paper re-derives the conditions under which the rational blockchain censorship attack of Yeo and Zhang behaves differently once validators are split into a small set of high-power "whales" and many low-power "minnows," showing that a detectability cost can make the attack unprofitable for whales and therefore hard to launch.

## What This Paper Is About

Earlier work (Yeo and Zhang, cited as [21]) showed that when validators arrive in random order and can coordinate through a single message on the blockchain, joining a censorship coalition is a dominant strategy and the attack is a subgame perfect equilibrium. This paper asks what happens under a more realistic voting-power distribution in which a few whales hold much more stake than many minnows: if the attack risks being detected and penalised, does joining still pay for the whales? The goal is to identify precise parameter regimes in which whales earn strictly less than their "fair play" reward by attacking, which would break the equilibrium and protect chain liveness.

## Key Contributions

1. **A whale/minnow model.** The authors instantiate the [21] attack with two validator classes — whales of voting power α and minnows of voting power ε — under the scalings w = O(1) and α/ε = O(n), where w is the number of whales and n the number of validators.
2. **An explicit profit analysis.** They decompose the whales' expected gain Φ(k,t) into a successful-attack term Φ₁(k,t) and a failed-attack term Φ₂(k,t), derive a hypergeometric upper bound Φ̄(k,t), and give a sufficient condition (Inequality 3) under which Φ̄(k,t) < αw, the whales' honest reward.
3. **A main theorem with an explicit threshold.** Theorem 4.1 gives a computable n₀(α, w, t) such that for all n ≥ n₀, some feasible m exists making Φ(k,t) < αw for all k in max{k_m, k̂(m,n), w} ≤ k < n.
4. **Empirical validation on real stake data.** Monte Carlo simulation on Ethereum and Solana stake distributions, fitted against exponential, power-law and stretched-exponential families, confirms the predicted reduction in whale reward.

## Main Findings

- **Attacking is not always profitable for whales.** The central result (Theorem 4.1) shows that once the detectability cost can actually be incurred — unlike in [21], where detection is effectively "in-built" and the attack either stops in time or goes undetected — there are parameter ranges where a whale's expected gain Φ(k,t) is strictly below its fair play value αw.
- **The whales' fair play value is αw.** Under the honest strategy the w whales collectively earn αw; any attack gain must be measured against this benchmark, and whales left outside the coalition earn nothing.
- **Attack success reduces to a whale-count condition.** Since S_k = αN_k + ε(k − N_k), the event {S_k ≥ t} is exactly equivalent to {N_k ≥ ⌈(t − kε)/(α − ε)⌉}, so the attack's viability is governed by how many whales appear among the first k arrivals.
- **Whale arrivals are hypergeometric.** N_k ~ Hypergeometric(n, w, k) with E[N_k] = kw/n and Var[N_k] = kw(n−k)(n−w) / (n²(n−1)); the variance term drives the concavity argument that yields the profit reduction.
- **A clean sufficient condition exists.** Because g(j,k) = αj/(αj + ε(k−j)) is strongly concave in j, Jensen's inequality gives a gain loss of α(α−ε)εk²w(n−k)(n−w)/(n²(n−1)), which must dominate the tail correction term to beat αw.
- **The feasible set for m is non-empty but constrained.** F = ℕ ∩ [ (t(n−w) − (1−αw)(n−1))/(αn−1), w−1 ], and Proposition 3 shows it is non-empty under the assumption α ≤ 1−t, i.e., w α < 1.
- **A closed-form threshold.** n₀(α, w, t) = ⌈ w + 1 + 1/α + (w−1)(2^w − w − 1) / (α(1 − αw)wt) ⌉.
- **Real chains show three regimes.** For t = 1/2, all four Monte Carlo curves (Ethereum and Solana, each with real whales' voting power and with equal whale power) show a flat phase at small k where attacking favours whales, a transition phase where attacking becomes detrimental, and a final upward phase that returns to αw — because at k = n everyone is included and the attack has no effect.
- **Stake distributions are heavy-tailed in both chains.** Stretched exponential fits best (R² = 0.994 Ethereum, 0.984 Solana), ahead of exponential (0.971, 0.947) and power law (0.927, 0.892), which justifies the whale/minnow split and supplies w = ⌈λ⌉ from the fitted characteristic length.
- **Concentration is real.** In Ethereum, Lido alone holds 22.3% of stake, followed by Binance at 8.3%; in Solana the top holder (Galaxy) holds 3.23%. This is the empirical basis for the whale assumption and is consistent with prior findings of stake concentration cited as [10], [15] and [19].

## Methodology in Plain English

The authors keep the attack protocol of [21] intact: validators arrive in a uniformly random order (a permutation π of the n validators), and a coalition stops growing either when its accumulated voting power S_k crosses the power threshold t, or when the coalition size reaches the detectability limit. The detectability parameter η from [21] is re-expressed here as (n − k)/n, and the detection cost is assumed to be considerably larger than any expected attack reward.

Onto this they graft a two-class power distribution: w whales each holding α, and n − w minnows each holding ε, with minnows' power normalised so total power is 1. Because attack success depends only on the number of whales N_k among the first k arrivals, and N_k is hypergeometric under a random arrival order, all relevant quantities can be written as finite sums over hypergeometric probabilities.

They then bound the whales' expected reward from both sides of the attack outcome: an exact expression for the case where the coalition crosses the power threshold at exactly k (Lemma 1), and an upper bound for the case where the coalition must wait longer and passes t later (Lemma 2). Trading the exact N₀ for a k-independent bound m gives the "hypergeometric upper bound" Φ̄ of Proposition 1. A tail inequality (Lemma 3) and a strong-concavity argument via Jensen's inequality then yield a sufficient condition for Φ̄ < αw, which Lemma 4 converts into a unique root k̂, and Lemma 5 shows the required quantity G(m,n) is monotone in m and n, delivering the closed-form n₀ in Theorem 4.1.

For the empirical part, they take validator counts and stakes from the source cited as [9], note that Ethereum's roughly 1M validators each hold 32 ETH, and therefore group validators into whales using the published stake-share percentages. This yields 10,040 validator units for Ethereum and 1,397 for Solana. They fit three heavy-tail families and use the stretched exponential's characteristic length λ, rounded up, as w. Finally they estimate Φ(k,t) by Monte Carlo at t = 1/2, once with real whale voting powers and once assuming all whales have equal power.

## Why This Matters

**Impact on research.** The paper does not overturn [21]; it adds the missing step that [21] left implicit — that the detectability cost must actually be chargeable for it to deter anyone. By showing that a plausible, empirically observed stake distribution (whales plus minnows) can flip the sign of whales' expected profit, it converts "censorship is a costless equilibrium" into "censorship is an equilibrium only outside a computable parameter window," and it gives that window in closed form. It also connects the game-theoretic censorship literature to the empirical stake-concentration literature.

**Real-world applications (bullets):**

- **Proof-of-stake protocol parameter setting:** the n₀ and feasible-set expressions give chain designers a way to reason about validator-set size and stake caps relative to a chosen detection threshold.
- **Staking-pool and liquid-staking risk assessment:** the Ethereum data shows the largest single holder at 22.3% of stake, so the "whale" branch of the model maps directly onto pool concentration risk.
- **Detection and monitoring policy:** the result that only detected attacks impose cost implies value in surveillance, attestation and slashing-style mechanisms that make large exclusions visible.
- **Validator-economics modelling:** the three-regime shape of the whale-reward curve is a usable design intuition for how coalition size interacts with reward dilution at different k.

**Industry relevance.** The results are directly relevant to staking providers, foundations overseeing proof-of-stake networks, and auditors who need to know whether a given stake distribution leaves the chain vulnerable to a self-organising censorship coalition. The Ethereum and Solana numbers (10,040 and 1,397 validator units; top shares of 22.3% and 3.23%) make the analysis concrete rather than hypothetical.

## Future Directions

- **Open problems section.** The paper promises a discussion of open problems in Section 6, but that section is not present in the supplied content, so its specific questions are not reported here. The natural gaps the text itself leaves are listed below.
- **Tightening the bound.** Theorem 4.1 bounds Φ through Φ̄, an upper bound on the whales' gain; the authors explicitly note that the guarantee holds "assuming the upper bound on the whale's gain is tight enough." Quantifying the gap between Φ and Φ̄ is an open step.
- **Beyond two classes.** The model assumes exactly two power levels with w = O(1) and α/ε = O(n); the text notes that for α/ε = Θ(n²) minnows collectively hold o(1) power and "being fair or unfair ceases to be a question," so intermediate and multi-tier distributions remain to be analysed.
- **Detection cost as a modelling choice.** The detection cost is a fixed drop in the cryptocurrency price assumed to dwarf attack rewards; how this cost is actually triggered, measured, and priced by markets is left unmodelled.
- **Extending the empirical study.** Thresholds other than t = 1/2, and chains beyond Ethereum and Solana, are not covered; the text also reports no numeric magnitude for the whale-reward reduction in Figure 2, only the qualitative three-regime shape, so quantifying that drop is a remaining task.

## Target Audience

Protocol researchers and security analysts working on proof-of-stake consensus and censorship resistance; game theorists and mechanism designers interested in validator incentives; and technically sophisticated readers at staking providers, foundations, or auditing firms who need to assess concentration risk in Ethereum- and Solana-class networks. The paper presumes comfort with probability bounds and equilibrium arguments, so it is not entry-level reading.

## Authors’ abstract

We study the blockchain censorship attack introduced in [21], which shows that joining the attack is a dominant strategy. We show that, by introducing certain detectability threshold, joining the attack can lead to strictly less reward for whales, which are defined to be a small number of validators that hold significantly more voting power than the rest (henceforth known as minnows). This leads to a change of the equilibrium: With whales unwilling to participate in the…

[The full abstract on arXiv](https://arxiv.org/abs/2608.15640)

[Read the original paper](https://arxiv.org/abs/2608.15640)

---

Canonical: https://crypto.world/paper/a-contribution-to-the-critique-of-blockchain-censorship-2608.15640
