This paper presents a novel Proof-of-Work (PoW) cryptocurrency protocol that addresses key limitations of Bitcoin and its recent variant, Tailstorm. The core innovation lies in combining Parallel Proof-of-Work (PPoW) consensus with DAG-style voting and a targeted reward discounting scheme. The protocol aims to provide superior consistency guarantees, higher transaction throughput, lower confirmation latency, and enhanced resilience against incentive-based attacks, such as selfish mining.
The work is motivated by the circular dependency in PoW systems between consensus algorithms and incentive schemes. While Bitcoin's properties are well-understood, many newer protocols lack thorough analysis of both consistency and incentives. Tailstorm improved upon Bitcoin but had shortcomings: its tree-structured voting left some votes unconfirmed, and its uniform reward discounting punished innocent miners alongside offenders.
PPoW, as introduced in prior work, requires a configurable number $k$ of PoW "votes" (or blocks) to be mined before the next main block can be appended. This creates a parallelized block structure. Each vote contains transactions. This design inherently provides stronger consistency guarantees than Bitcoin's linear chain because finalizing a block requires multiple supporting proofs.
Tailstorm structured these $k$ votes as a tree, where each new vote references a single parent. This creates a dilemma: miners must choose which branch to extend, leaving some branches—and their transactions—unconfirmed until the next block.
The proposed protocol structures votes as a Directed Acyclic Graph (DAG). A new vote can reference multiple previous votes as parents. This increases connectivity and allows more votes to be included in the consensus set for a given block, improving transaction confirmation rates and reducing latency.
Tailstorm discounted rewards proportionally to the depth of the vote tree, punishing all miners in a deep (non-linear) tree equally. The new protocol implements a targeted discounting scheme. The reward for a miner's vote is calculated based on its specific role in the DAG:
$Reward_v = BaseReward \times (1 - \alpha \cdot C_v)$
Where $C_v$ is a measure of the vote $v$'s contribution to non-linearity or fork creation (e.g., how many competing votes it references that are not themselves connected). The parameter $\alpha$ controls the discount strength. This ensures only miners whose actions directly harm consensus linearity are penalized.
The paper claims that after a 10-minute confirmation window, the probability of a successful double-spend attack is approximately 50 times lower than in Bitcoin, under realistic network assumptions. This stems from the $k$-vote requirement in PPoW, which makes it statistically harder for an attacker to reverse a confirmed block.
A significant methodological contribution is the use of Reinforcement Learning (RL) to systematically search for optimal attack strategies against the protocol. The RL agent learns to manipulate vote publication timing and parent selection to maximize profit. This approach is more rigorous than ad-hoc attack analysis and revealed that vanilla PPoW (without discounting) is vulnerable.
The combination of DAG voting and targeted discounting creates a powerful disincentive for selfish mining. Attacks that involve withholding blocks or creating forks become less profitable because the attacker's rewards are directly discounted. The RL-based analysis confirms the proposed protocol's superior resilience compared to both Bitcoin and Tailstorm.
By packing transactions into each of the $k$ votes per block, the protocol achieves higher throughput than Bitcoin's single-block-per-interval model. The DAG structure further reduces latency by allowing more votes (and thus their transactions) to be confirmed in the current block rather than being deferred.
The paper directly addresses Tailstorm's two flaws: 1) Unconfirmed Votes: DAG mitigates this by allowing multiple parent references. 2) Collective Punishment: Targeted discounting replaces uniform tree-depth punishment. The result is a protocol that retains Tailstorm's benefits while overcoming its weaknesses.
The reward discounting function is central. Let $G$ be the DAG of votes for a block. For a vote $v \in G$, define its "conflict score" $C_v$. One proposed measure is:
$C_v = \frac{|\text{Unconnected Parents}(v)|}{|\text{Total Parents}(v)| + \epsilon}$
Where "Unconnected Parents" are parent votes that are not themselves ancestrally linked. A high $C_v$ indicates $v$ is referencing conflicting branches, increasing non-linearity. The final reward is discounted by this score. The RL agent's objective is to learn a policy $\pi$ that maximizes cumulative discounted reward $\sum \gamma^t R_t$, where $R_t$ is the (potentially discounted) reward from publishing a vote at time $t$ with specific parent selections.
The paper likely includes simulations comparing attack success rates and profitability across Bitcoin, Tailstorm, basic PPoW, and the proposed DAG-PPoW with targeted discounting. Key expected results presented in charts or tables would show:
Critical Finding: The experiments presumably confirm the paper's striking claim that "parallel proof-of-work without reward discounting is less resilient to incentive attacks than Bitcoin in some realistic network scenarios." This underscores the absolute necessity of coupling new consensus mechanisms with carefully designed incentive schemes.
Scenario: A miner (M) controls 25% of the network hash rate and wants to execute a selfish mining attack.
In Bitcoin/Tailstorm: M withholds a found block to create a private fork. If successful, M can orphan honest blocks and claim a disproportionate reward. The RL agent would learn this strategy.
In DAG-PPoW with Targeted Discounting:
This case shows how the protocol's mechanics directly alter the attacker's profit calculus.
This paper isn't just another incremental tweak on Proof-of-Work; it's a surgical strike on the fundamental incentive-consensus loop that plagues blockchain design. The authors correctly identify that most "improved" protocols fail because they optimize for liveness or throughput in a vacuum, ignoring how those changes warp miner economics. Their key insight is that security isn't a property of the consensus algorithm alone, but of its tight coupling with a penalty system that can precisely attribute blame. Moving from Tailstorm's tree to a DAG isn't about efficiency—it's about creating the forensic granularity needed for targeted punishment.
The argument builds impeccably: 1) Bitcoin's limits are well-known, 2) Tailstorm made progress but introduced new problems (blunt punishment, deferred confirmations), 3) Therefore, we need a structure (DAG) that provides finer-grained data on miner behavior, and 4) We must use that data to enact surgical disincentives. The use of Reinforcement Learning to stress-test the proposal is particularly elegant. It mirrors how real-world attackers operate—not following static scripts, but adaptively searching for profit—and thus provides a more realistic security assessment than traditional probabilistic models. The shocking finding that vanilla PPoW can be less secure than Bitcoin is a testament to this method's value; it exposes hidden attack surfaces.
Strengths: The conceptual framework is robust. The DAG+targeted discounting mechanism is elegant and addresses clear flaws in prior art. The methodological rigor (RL-based attack search) sets a new standard for evaluating crypto-economics. The paper also usefully demystifies the often-overhyped "DAG" term, applying it to a specific, measurable purpose within a PoW context, unlike more speculative DAG-based projects.
Flaws & Open Questions: The elephant in the room is complexity. The protocol requires miners and nodes to maintain and analyze a DAG, compute conflict scores, and apply custom discounts. This increases computational and implementation overhead compared to Bitcoin's beautiful simplicity. There's also a risk of the discounting parameters ($\alpha$) becoming a source of governance conflict. Furthermore, as with many academic proposals, the analysis likely assumes a somewhat rational, profit-maximizing miner. It doesn't fully address Byzantine actors whose goal is disruption rather than profit—a threat model considered in traditional BFT literature like that of Castro and Liskov (1999).
For protocol designers: Incentive analysis is non-negotiable. Any consensus change must be modeled with tools like RL to uncover perverse incentives. The "PPoW-less-secure-than-Bitcoin" finding should be a wake-up call. For developers: The DAG-for-accountability pattern is a powerful tool worth exploring in other consensus contexts, perhaps even in sharded architectures or layer-2 networks. For the research community: This work highlights the urgent need for standardized, open-source RL frameworks for attacking crypto-economics, similar to how the AI community has benchmark datasets. Finally, the biggest takeaway is that blockchain security is moving from pure cryptography to a hybrid discipline of cryptography, game theory, and machine learning. Future secure systems will need expertise in all three.