Research
Temporary Power Adjusting Withholding Attack
In a block withholding attack, an adversary joins a mining pool, submits partial proofs of work (pPoWs) to look productive, but hides the full proofs of work (fPoWs) it finds, while still collecting…

- arXiv
- 2604.14135
- Published
- 2026-04-15
- Authors
- Mustafa Doger, Sennur Ulukus
AI summary
Overview
Research area: Blockchain security, specifically Proof-of-Work (PoW) mining-pool reward manipulation and stochastic modeling of adversarial mining strategies.
Technical level: Advanced. The paper is built on continuous-time stochastic process analysis (exponential block arrivals, incomplete gamma / exponential-integral functions), reward-ratio optimization over three parameters, and familiarity with prior withholding attacks (BWH, FAW, PAW).
Scope (one sentence): The paper generalizes the Power Adjusting Withholding (PAW) pool attack by bounding the withholding time to a finite value (T), derives exact revenue expressions, and shows through optimization that the bounded version earns strictly more than PAW for a wide range of parameters.
What This Paper Is About
In a block withholding attack, an adversary joins a mining pool, submits partial proofs of work (pPoWs) to look productive, but hides the full proofs of work (fPoWs) it finds, while still collecting a share of the rewards earned by the pool's honest members. PAW (Power Adjusting Withholding) refines this by re-allocating the adversary's hashpower toward the target pool once a fPoW is withheld, and it holds that fPoW indefinitely until someone else finds a block. This paper asks whether holding forever is actually the best choice. The authors propose Temporary PAW (T-PAW), in which the adversary releases the withheld fPoW after at most (T) time even if nobody else has mined a block, and they optimize (T) jointly with the two power-allocation fractions.
Key Contributions
-
A new attack, T-PAW, that nests existing attacks. The adversary withholds a pool-mined fPoW for at most (T) time. Setting (T=0) recovers honest mining regardless of the power allocations, and setting (T \to \infty) recovers PAW. The appendices further show how PAW, FAW, and BWH are recovered from T-PAW by specific choices of the fork-race parameter (c), (p_1), (p_2), and (T).
-
Proof that PAW is not optimal. The paper shows PAW corresponds to (T \to \infty) and is sub-optimal to T-PAW, with the improvement factor growing without bound as the adversarial hash fraction (\alpha), pool size (\beta), and adversarial network influence (\gamma) decrease.
-
An exact treatment of pool reward sharing. The authors replace an approximation used in prior PAW analyses (references [8] and [9]) regarding how fPoW rewards are split among pool members with an exact derivation based on the stochastic distributions of mining times. They state the approximation error grows large especially for T-PAW.
-
A profitability result that does not depend on difficulty adjustment. T-PAW is shown to dominate honest mining even with no difficulty adjustment algorithm (DAA), and to produce no profit lag in most parameter regimes even when the objective is revenue-ratio maximization.
Main Findings
-
Unbounded improvement in the small-parameter regime. The extra reward of T-PAW over PAW improves by an unbounded factor as (\alpha), (\beta), and (\gamma) decrease. At ((\alpha,\beta,\gamma)=(0.05,0.05,0)), the extra reward of T-PAW is 22 times that of PAW. At (\gamma=0), (\alpha=0.06), (\beta=0.06), the relative extra reward of T-PAW is 15 times that of PAW.
-
Non-trivial short-term gains without any DAA. For most ((\alpha,\beta)), the adversary gains at least 1% within two weeks in Bitcoin even when (\gamma=0), whereas for PAW the corresponding gain was at most 0.01%. In the revenue-change maximization at (t_1) with (\gamma=0), T-PAW always yields a positive revenue change of more than 0.1%, and inside the marked region the gain exceeds 1%; the PAW case only has a no-profit-lag region below a boundary line, with less than 0.01% gain.
-
The adversary, not the outsiders, benefits most. Prior work showed that in PAW honest miners outside the target pool often profit more than the attacker, especially at low (\gamma). In T-PAW, even when (\gamma=0), the adversary benefits most from the attack. The paper also proves the general bound (\rho_{rest} \ge 1-\alpha-\beta), so non-pool miners never lose revenue share.
-
A structural boundary in the optimal strategy. Numerical optimization reveals a diagonal boundary. Above it, (\lambda_1 T^{\ddagger} \approx \infty) and (p_1^{\dagger}=p_1^{\ddagger}), (p_2^{\dagger}=p_2^{\ddagger}), meaning optimal T-PAW collapses to PAW. Below it, (\lambda_1 T^{\ddagger} < 1) and (p_2^{\ddagger}=1): the adversary withholds only briefly and fully commits its power to the pool, because it knows the fPoW will be shared soon. Inside that region (\lambda_1 T^{\ddagger}) increases monotonically with (\beta).
-
Intuition for why finite withholding wins. Exponential arrivals are memoryless, so the time until someone else mines does not depend on how long the fPoW has been held. But the adversary's reward from that fPoW is not memoryless: its fractional contribution saturates over time while its exposure to losing the fork race keeps growing, particularly when (\gamma) is small.
-
Fork-race influence drives the optimal (T). Across 5 tested ((\alpha,\beta)) cases, optimization gives (p_2^{\ddagger}=1) for all (\gamma), whereas PAW's (p_2^{\dagger}=1) only at high (\gamma). (\lambda_1 T^{\ddagger}) increases monotonically with (\gamma), and as (\gamma \to 1), T-PAW reduces to PAW. Small ((\alpha,\beta)) with low (\gamma) gives more than a 20-fold improvement in relative extra reward.
-
Only the product (\lambda_1 T) matters. Because all expected reward quantities depend on (\lambda_1 T), scaling the block rate can be absorbed by rescaling (T), which avoids re-solving the optimization after a difficulty adjustment.
Methodology in Plain English
The authors model the blockchain as a system with a large number of miners, each with a negligible share, where block arrivals follow an exponential distribution with total rate (\lambda_1). A fraction (\alpha) of hashpower is adversarial and colluding; a target pool of honest miners holds fraction (\beta); the analysis assumes (\alpha+\beta<0.5). The adversary splits its power, putting fraction (p_1) into pool mining while waiting for a fPoW, then switching to fraction (p_2) during the withholding phase.
They analyze one "attack cycle" at a time. A cycle can end in five ways: an honest outsider mines a block, the adversary mines individually, a pool member mines, the withheld fPoW triggers a fork race when an outsider mines, or the timer (T) expires and the adversary releases the fPoW anyway. Because cycles are independent, the long-run revenue ratio equals the expected adversarial reward per cycle divided by the expected number of canonical blocks per cycle, and the authors derive closed-form expressions for these expectations, including terms involving the exponential integral (E_1(\cdot)) for the case where the timer expires.
They define three metrics: Relative Extra Rewards (RER), which compares revenue ratios under two strategies; Revenue Change over time, which accounts for block redundancy and difficulty epochs of (D_0) canonical blocks with readjustment (d_n = d_c \tau_0/\tau_c); and Profit Lag, the earliest time after which the strategy always beats honest mining. Since the objective functions are neither convex nor concave in general, they use MATLAB's fmincon from at least 100 different initial values of ((p_1,p_2,T)), take the best local maximum, and sweep (\alpha) and (\beta) in increments of (\epsilon = 0.003) for (\gamma \in {0, 0.5}), plus (\gamma) ranging over ([0,1]) for 5 selected ((\alpha,\beta)) pairs.
Why This Matters
The attack reframes pool mining's security story. Instead of a large attacker exploiting a pool, the paper argues that the pool's own primary participants — small home miners — can individually turn into profitable adversaries with immediate, non-trivial gains. That is a structural weakness in pooled mining rather than a niche parameter corner.
Real-world applications and implications:
- Pool payout and share-accounting design: pools that weight rewards by submitted pPoWs are directly exposed, and the analysis shows exactly how a member's optimal withholding window scales with block rate.
- Countermeasure evaluation: the paper argues existing defenses are weaker against T-PAW. The timestamp-field proposal of [8] is less effective because the adversary gains even within a limited (T); silent time-stamping [20] can be circumvented by choosing nonce step sizes differently before and during withholding, or by using Sybil identities; the anti-withholding reward system of [19] (paying more for fPoW than pPoW) helps against BWH, FAW, and PAW mainly at low (\gamma) but not against T-PAW, since the adversary usually releases its fPoW before anyone else does.
- Difficulty adjustment and protocol research: T-PAW is profitable with no DAA at all, which changes the baseline assumption used to argue that withholding strategies are unprofitable absent difficulty manipulation.
- Miner economics and pool choice: small miners can evaluate whether temporary withholding beats honest participation for their own ((\alpha,\beta,\gamma)) region, and pool operators can identify which pool sizes are most vulnerable.
Industry relevance: the results apply to any PoW chain using pooled mining with pPoW-based share accounting and epoch-based difficulty retargeting, with Bitcoin used as the concrete reference for the two-week, 1%-gain claim; the historical Eligius incident (300 BTC lost in 2014 after a withholding attack was announced) is cited as precedent that these attacks are not purely theoretical.
Future Directions
- Multi-pool analysis. The paper identifies extending the exact T-PAW model to the competing-pools "prisoner's dilemma" setting studied for FAW/PAW in [7] and [8] as "the most natural and promising future direction," including how T-PAW shifts the parameter regimes where the dilemma holds.
- Rigorous countermeasure analysis. The discussion of timestamping, silent time-stamping, Sybil identities, and anti-withholding reward schemes is qualitative; a detailed analysis against each countermeasure is explicitly left to future work.
- Combining with other attack vectors. The authors note that eclipse attacks [22], bribery [8, 23, 24], Uncle Block Attack stale-block rewards [21], and high-propagation-delay models [25] could be jointly considered to extend scope and applicability.
- Non-rational pool managers and richer payout schemes. The paper assumes a rational pool manager ((\mathbb{1}_{P_R}=1)) and notes the improvements over PAW become even stronger otherwise; share-withholding manipulation of pPoW timing under time-decaying payouts [18] is another open combination, as is handling non-convex objectives more reliably.
Target Audience
Blockchain security researchers and graduate students working on consensus-layer incentive analysis; mining-pool operators and payout-scheme designers; protocol engineers responsible for difficulty adjustment and stale-block policy; and quantitatively inclined readers who already understand Proof-of-Work, pooled mining, and basic probability but want a rigorous, self-contained treatment of how far withholding attacks can be pushed.
Authors’ abstract
We consider the block withholding attacks on pools, more specifically the state-of-the-art Power Adjusting Withholding (PAW) attack. We propose a generalization called Temporary PAW (T-PAW) where the adversary withholds a fPoW from pool mining at most $T$-time even when no other block is mined. We show that PAW attack corresponds to $T\to\infty$ and is not optimal. In fact, the extra reward of T-PAW compared to PAW improves by an unbounded factor as adversarial hash fraction…
The full abstract on arXiv