Skip to content
Crypto.world

Research

On the Effectiveness of Mempool-based Transaction Auditing

Blockchains confirm transactions in blocks, but before confirmation a miner can silently drop, delay, or reorder the transactions sitting in its mempool — censoring or "displacing" them for profit.

On the Effectiveness of Mempool-based Transaction Auditing
arXiv
2601.14996
Published
2026-01-21
Authors
Jannik Albrecht, Ghassan Karame

AI summary

On the Effectiveness of Mempool-based Transaction Auditing

Overview

  • Research area: Blockchain security and peer-to-peer network measurement — specifically transaction-ordering attacks (censorship, front-running, transaction displacement) and the informal "mempool auditing" tools used to detect them in Bitcoin and Ethereum.
  • Technical level: Advanced. The paper builds formal probability models (integrals over propagation-time distributions, binomial/conditional probability theorems) and validates them with real-world node measurements. The prose is readable, but the analysis sections assume comfort with probability theory.
  • Scope (one sentence): The paper quantifies, both analytically and empirically, how reliably observer-based mempool auditing can detect miner misbehavior, and what that implies for fair-ordering schemes.

What This Paper Is About

Blockchains confirm transactions in blocks, but before confirmation a miner can silently drop, delay, or reorder the transactions sitting in its mempool — censoring or "displacing" them for profit. Since no clean-slate protocol-level defense has been adopted by Bitcoin, Ethereum, or Cardano, the community instead relies on ad-hoc tools like Mempool.space, which compare what independent "observer" nodes saw with what miners actually put into blocks. This paper asks how much those audits can actually be trusted, and how often they accuse honest miners of wrongdoing.

Key Contributions

  1. A realistic network model for mempool auditing. The authors define a model with n observers, a propagation-time distribution D(r) for transactions, and a public deterministic ordering algorithm, then derive the probability that a single observer receives transaction t1 before t2 sent ω time later (Equation 1).
  2. Two central theorems on observer agreement and inference. Theorem 1 (Q_n(m)) gives the probability that exactly m of n observers see the same order; Theorem 2 (P_n(m)) gives the probability that t1 really was sent before t2, conditioned on m of n observers reporting that order — i.e., the precision of an audit.
  3. Large-scale empirical validation. Sixteen geographically dispersed Bitcoin and Ethereum nodes were run over 196 hours (30.06.–08.07.2025), recording 3,134,824 Bitcoin transactions, used to check the analytical model against reality.
  4. First analysis of batch-order fair ordering in practice. The results are applied to threshold fair-ordering schemes such as Themis (Kelkar et al., 2023) and compared with traditional fair-ordering approaches Pompe (Zhang et al., 2020) and Wendy (Kursawe, 2020).

Main Findings

  • False accusations are common. Mempool auditing can mislabel honest miners as "malicious" with a probability larger than 25% in some settings (Section 4.2).
  • 30 seconds is the decisive threshold. Auditing succeeds with an overwhelming probability of 99.9% when two transactions are consistently received by all observers and sent at least 30 seconds apart.
  • Short gaps destroy unanimity. When two transactions are sent ω ≤ 5 seconds apart, the chance that all m = 5 observers see the correct order is at most 23%; only at ω ≥ 60 seconds does it exceed 99%.
  • Large observer sets cut both ways. With n = 16 observers, transactions sent ω ≤ 30 seconds apart are received in the same order by all 16 with probability less than 78% — i.e., 22% of such pairs arrive inconsistently. Formally, n1 > n2 ⇒ Q_{n1}(n1) ≤ Q_{n2}(n2).
  • Majority thresholds recover reliability. Receiving transactions with ω ≥ 30 in order is nearly guaranteed for 4 out of 5 observers and for 12 out of 16 observers. Precision values from the KIT-based evaluation include P_5(5) = 99.90%, P_12(12) = 100.0%, and P_16(16) = 100.00%.
  • More observers is not better. The paper finds the ideal number of observers for precise and reliable observations to be small — roughly two to six — in contrast to the n = 24 observers used by Mempool.space.
  • Real-world Bitcoin data corroborates the model. Only 93.93% of transactions were received by every observer; 99.08% reached at least m ≥ 12 nodes and 99.34% reached at least m ≥ 9. Each node averaged 98.9% ± 0.15% of recorded transactions, and the average pairwise Mutual Information between observers was 0.15 nats, supporting the independence assumption.
  • Empirical ordering agreement is high but not total. Of 100,000 randomly sampled transaction pairs (200,000 transactions), 88.23% were witnessed in the same order by all 16 observers when drawn from the complete dataset, rising to 99.98% when restricted to transactions fully received by all 16 observers.
  • Fair-ordering guarantees are narrow. As a direct consequence, batch-order fair-ordering schemes can offer strong fairness guarantees only for a limited subset of transactions in real-world deployments, because their core assumption — that validators receive a consistent transaction order — is rarely satisfied.

Methodology in Plain English

The authors treat transaction delivery as a race against the clock. Each observer receives a transaction after some random delay; if two transactions are sent close together, jitter can flip their apparent order at any given node. Using a published measurement of Bitcoin propagation delays (the KIT dataset from Neudecker et al., 2016), they fit a log-normal distribution with mean μ = 1.973 and variance σ² = 0.585 and use it to compute the odds that a given number of observers will agree on an order. They report that 90% of transactions are received after 20 seconds and 99% after at most 27.75 seconds, with a long tail beyond 30 seconds.

Two questions follow. First, how likely is unanimity among observers? Second, given that m of n observers agree, how likely is it that the miners actually misbehaved rather than that the network simply jittered? The second question is answered with Bayes-style conditioning, which yields the audit's precision. For censorship across blocks, a separate theorem computes the probability that a whole sequence of later transactions overtakes an earlier one before a block-filling cutoff time τ_C, namely (1 − F(τ_C))^n.

To check the math, the team deployed 16 Bitcoin and Ethereum nodes on Hetzner cloud servers (four vCPU, 16GB RAM, SSD, AMD EPYC-Milan, Ubuntu 24.04 LTS) using bitcoind 25.0 and geth 1.15.11, spread across the globe, and logged every incoming transaction with an exact timestamp for 196 hours. They then sampled 100,000 transaction pairs and measured how often observers agreed on the order, comparing this against the model's predictions.

Why This Matters

Mempool auditing is not a theoretical proposal — it is already used by public block explorers to rank mining pools by "block health," and that ranking can shape user and pool reputation. The paper shows this deterrent mechanism can misfire on honest miners, and it supplies the first quantitative bound on when an audit result is trustworthy.

Real-world applications:

  • Block explorer operators (Mempool.space, Litecoin.space, Blockstream Explorer, Etherscan and similar) can pick observer counts and time thresholds that match the paper's precision guarantees instead of assuming more observers is strictly better.
  • Mining pools gain a defensible standard for contesting accusations of censorship or displacement, based on whether the flagged transactions were sent at least 30 seconds apart and consistently received.
  • Users and trading firms relying on "block health" scores to choose where to route transactions can calibrate how much weight those scores deserve.
  • Protocol designers working on fair ordering (Themis, Pompe, Wendy) get an empirical constraint showing that consistent-ordering assumptions hold for far fewer transactions than assumed.

Industry relevance: The findings directly challenge the operating assumptions of deployed transparency tooling and of the fair-ordering research agenda, and they identify a concrete, measurable separation (roughly 30 seconds) below which ordering evidence is not actionable. Qin et al.'s reported losses of $540.54M over 32 months (Dec 2018–Aug 2021) from arbitrage attacks illustrate the financial stakes behind getting detection right.

Future Directions

  • Designing audit thresholds that bound false accusations, rather than maximizing recall, including optimal observer counts in the two-to-six range the analysis favors over the 24 used by Mempool.space.
  • Fair-ordering schemes that tolerate inconsistent inputs. Since validators rarely share a consistent transaction order, follow-up work must define fairness guarantees for the subset of transactions that pass the 30-second separation test, or design schemes robust to the rest.
  • Extending the analysis beyond Bitcoin to Ethereum and other chains, and testing whether the mixture of block building under proposer-builder separation (PBS) introduces additional ordering inconsistencies, as the paper's footnote flags.
  • Re-measuring with alternative propagation models and topologies, since the analytical results are parameterized by an assumed propagation distribution D(r) and validated against a single 196-hour measurement window in 2025.

Target Audience

Blockchain security researchers and graduate students studying transaction-ordering attacks and MEV; engineers and operators of block explorers and mempool-monitoring services; mining pool operators concerned with reputation metrics; and protocol designers evaluating fair-ordering schemes. Readers need a solid probability background to follow the theorem sections, but the empirical results and conclusions are accessible to intermediate-level practitioners.

Authors’ abstract

While the literature features a number of proposals to defend against transaction manipulation attacks, existing proposals are still not integrated within large blockchains, such as Bitcoin, Ethereum, and Cardano. Instead, the user community opted to rely on more practical but ad-hoc solutions (such as Mempool.space) that aim at detecting censorship and transaction displacement attacks by auditing discrepancies in the mempools of so-called observers. In this paper, we…

The full abstract on arXiv
Read the original paper