# All-out Attack: Optimal Block Withholding Under Pay-Per-Share Scheme

Crypto.world

> Most prior block-withholding research assumes block-dependent payouts (PROP, PPLNS), where pool members are paid only when the pool actually finds a block.

**arXiv**: 2607.01209

**Published**: 2026-07-01

**Authors**: Mustafa Doger, Sennur Ulukus

## AI summary

## Overview

**Research area:** Blockchain security and mechanism design — specifically incentive attacks on Nakamoto-style Proof-of-Work mining pools, bridging classical block-withholding (BWH) literature with the share-based payout schemes that dominate modern mining.

**Technical level:** Intermediate. The paper uses revenue-comparison analysis and light probability modelling rather than heavy formalism; the core argument (difficulty adjustment changes the economics) is accessible to anyone who understands how mining pools and difficulty retargeting work.

**Scope (1 sentence):** The paper proves that under Pay-Per-Share (PPS) payouts, withholding every full Proof-of-Work block while submitting only partial-Proof-of-Work shares — the "All-out Attack" — is the profit-maximizing strategy once difficulty adjustment is modelled, contradicting prior claims that PPS is incentive compatible.

## What This Paper Is About

Most prior block-withholding research assumes block-dependent payouts (PROP, PPLNS), where pool members are paid only when the pool actually finds a block. Most real pools today instead use share-based payouts such as PPS and FPPS, paying miners immediately for every valid share. Earlier work — notably Rosenfeld's sabotage analysis — concluded that withholding blocks under PPS yields the attacker no profit and is merely vandalism, but those conclusions assumed a fixed block difficulty.

This paper re-examines the question with difficulty adjustment (DA) in the model, and asks what the *optimal* hashpower allocation for a withholding attacker under PPS actually is. The answer is that the attacker should send its entire hashpower to the victim pool and never release a single block.

## Key Contributions

1. **Refutes the incentive-compatibility claim for PPS under difficulty adjustment.** The authors show that the earlier "no profit from withholding under PPS" results hold only with fixed difficulty, and that modelling the DA mechanism reverses the conclusion.
2. **Characterizes the optimal attack as the All-out Attack (AoA).** Theorem 1 proves that setting the pool-directed hashpower fraction to p* = 1 maximizes adversarial revenue change after the first difficulty adjustment under PPS, provided α > 2^{-Z}; otherwise honest mining is optimal.
3. **Derives the optimal allocation under variable share difficulty (vardiff).** Theorem 2 gives p* = min{1, 1/(α + 2q)} with q ≜ T_share/T_b, and Corollary 2 shows that with arbitrarily many sybil identities q can be driven arbitrarily small, making AoA optimal in general.
4. **Quantifies the revenue change of every stakeholder separately.** Unlike zero-sum revenue-ratio analysis, the authors track the adversary, honest pool members, external honest miners, and the pool operator, showing that all miners gain equally per unit hashpower while the pool operator absorbs the entire loss.

## Main Findings

- **AoA revenue rate:** After the first difficulty adjustment, the attacker achieves a reward rate of α/(1−α) per target epoch, versus an honest baseline rate of α.
- **Matches the selfish-mining ceiling:** This α/(1−α) gain is stated to match the theoretical upper bound of optimal selfish mining in Nakamoto consensus under perfect network influence (γ = 1), even though the underlying mechanism is different.
- **Negligible pre-adjustment loss:** In the first epoch the adversary's loss from withheld full-PoW blocks is −δ_p^BWH · αp · 2^{-Z} ≈ 0, since 2^{-Z} ≪ 1. The first epoch lasts t₁ = τ₀/(1 − αp).
- **Two channels for post-adjustment profit:** If the pool scales pPoW difficulty down along with fPoW difficulty, the attacker earns more shares per unit time; if share difficulty is held fixed, each share is worth more because PPS reward scales with the share-difficulty-to-block-difficulty ratio (R_s = R_b · 2^{-Z}).
- **Two break-even conditions:** Under fixed relative share difficulty, AoA is optimal iff α > 2^{-Z}; under vardiff, AoA is optimal iff α < 1 − 2q, and q < 1/4 suffices given the standing assumption α + β ≤ 1/2.
- **Practice satisfies the condition easily:** In Bitcoin, if a pool targets one share per second, q = 1/600 ≈ 1.67×10^{-3}, more than two orders of magnitude below the 1/4 threshold; even a one-minute share interval gives q = 60/600 = 0.1 < 1/4.
- **Honest miners free-ride:** Every other mining entity (honest pool members and external miners) gains at the same rate per unit hashpower as the attacker, roughly δ₁^BWH − 1 = α/(1−α) per epoch.
- **The pool operator bears everything:** The operator pays out-of-pocket for pPoW submissions it never recovers via fPoW, losing α · δ₁^BWH per epoch, and discovers blocks at only β/(α+β) of its expected rate — a deviation potentially detectable within an epoch.
- **Advanced BWH variants add nothing under PPS:** FAW, PAW and T-PAW yield no additional adversarial profit because releasing a block can raise difficulty and reduce mining efficiency; FAW can only marginally mitigate the operator's losses, and only if orphan blocks do not feed into the difficulty calculation.
- **Comparison with block-dependent schemes:** Even the state-of-the-art T-PAW attack against PROP with γ = 1 has a relative extra revenue (RER) strictly below that of AoA under PPS at every α (citing [11, Fig. 4]), where AoA reaches RER = α/(1−α).
- **Fees do not rescue the scheme:** With a fee where a miner receives rR_s per share (0 ≤ r ≤ 1), the post-DA rate becomes α(1 − cp)/(1 − αp) with c = 1 − r(1 − 2^{-Z}), and AoA remains optimal roughly whenever α exceeds the pool fee 1 − r.
- **Simulation agreement:** Monte Carlo runs with T_share = 1 s, T_b = 600 s, D₀ = 2016 blocks, N = 1000 independent runs and 99% confidence intervals closely match the theoretical predictions.

## Methodology in Plain English

The authors avoid the zero-sum "revenue ratio" framework normally used for block-dependent pools, because under PPS the operator's reserve capital enters the picture and one party's gain is not simply another's loss. Instead they use **revenue change analysis**: they pick a time t = 0 where the attacker starts misbehaving, compute what each party would have earned if everyone had been honest, and subtract.

The model is a longest-chain PoW network with n miners, an adversary holding α of the hashpower and a victim pool holding β, constrained by α + β ≤ 0.5. An epoch is defined as the time to accumulate D₀ canonical blocks, targeting τ₀ on average, with total block reward normalized to 1. After each epoch, a Bitcoin-style difficulty adjustment rescales block difficulty by τ₀/τ, where τ is the observed epoch duration.

A key bookkeeping quantity is the **block redundancy ratio** δ = 1/(1 − αp), the ratio of total blocks mined during an attack renewal period to canonical blocks; withholding αp of hashpower stretches each epoch by this factor. The analysis then splits rewards into two buckets — coinbase rewards actually earned, and PPS payments transferred from operator to members — and evaluates each for the adversary, honest pool members, external miners, and the operator. Two settings are treated: fixed relative share difficulty (constant Z = Z_f − Z_p) and practical vardiff, where the pool pins the share interval so that 2^{-Z} = αpq.

## Why This Matters

**Impact on research:** The paper overturns a widely cited conclusion (Rosenfeld's sabotage analysis and follow-ups) that PPS is incentive compatible with respect to withholding, by showing the result depends on the fixed-difficulty assumption. It also connects a pool-level attack to the selfish-mining upper bound, and it demonstrates a methodological point: zero-sum revenue-ratio analysis does not transfer to share-based schemes, so a different tool (revenue change analysis with the operator's reserves modelled) is required.

**Real-world applications:**

- **Pool operator risk management:** PPS/FPPS operators must hold large reserves; this work shows those reserves are the attack surface, and gives the per-epoch loss α · δ₁^BWH.
- **Detection:** The victim pool's block discovery rate falling to β/(α+β) of expectation is a statistically observable signature, though the paper notes that spreading the attack across pools dilutes it.
- **Payout mechanism design:** The result argues for payout designs that account for difficulty adjustment rather than treating share difficulty and block difficulty independently.
- **Miner and investor due diligence:** Since honest miners benefit at the same per-hashpower rate as the attacker, participants evaluating pool health should understand that the operator, not the membership, absorbs the damage.

**Industry relevance:** The paper notes that Foundry USA (approximately 25–28% of network hashrate, using FPPS), AntPool (approximately 19–21%, offering FPPS, PPS and PPLNS), F2Pool (approximately 12–14%, FPPS/PPS variants) and ViaBTC (approximately 10–11%, PPS+ and PPLNS) collectively account for a substantial fraction of Bitcoin's hashrate — i.e., the vulnerable payout class dominates the market. The historical precedent cited is Eligius, which announced in 2014 that it had lost 300 BTC to a block-withholding attack.

## Future Directions

- **Pool-side detection mechanisms:** The paper explicitly leaves designing detection to future work; the observable signal (block rate at β/(α+β) of expectation) and its dilution across multiple pools and multiple identities needs a concrete statistical test.
- **Payout designs robust to difficulty adjustment:** The authors call for reward schemes that internalize DA effects, since the α > c condition in the fee analysis shows a fee alone does not eliminate the incentive.
- **Faster difficulty adjustment algorithms:** Remark 3 conjectures the attack is *at least* as effective under faster DAAs, because profitability arrives immediately after the first downward adjustment and the low-loss transient is shorter — an untested claim.
- **Uncle-block and orphan-block protocols:** Remark 5 and the FAW discussion leave open how uncle-block difficulty accounting changes the trade-off, including protocols where releasing withheld blocks partially offsets operator losses.

## Target Audience

Mining pool operators and their risk/compliance teams; protocol and mechanism-design researchers working on incentive compatibility of PoW consensus; security analysts covering the mining industry; and graduate students studying selfish mining and block withholding, who will find a clean worked example of why fixed-difficulty assumptions can flip a security conclusion. Readers need basic familiarity with Proof-of-Work, mining pools and difficulty retargeting, but not with the prior BWH literature.

## Authors’ abstract

Classical Block Withholding (BWH) attacks have been extensively studied in block-dependent reward schemes, where pool members are compensated upon a block discovery within the pool. However, most contemporary mining pools operate under share-based schemes, wherein participants are paid immediately upon submission of valid shares. In this paper, we analyze BWH under Pay-Per-Share (PPS) for Nakamoto-style blockchains and prove that these mechanisms are not incentive…

[The full abstract on arXiv](https://arxiv.org/abs/2607.01209)

[Read the original paper](https://arxiv.org/abs/2607.01209)

---

Canonical: https://crypto.world/paper/all-out-attack-optimal-block-withholding-under-pay-per-share-scheme-2607.01209
