TL;DR
This paper provides a mathematical framework for designing NFT-based reward systems that are resilient to Sybil farming. By combining reward vesting, stochastic auditing, and cluster-level penalties, the authors demonstrate how to deter automated attacks while minimizing friction for genuine participants.
What Problem Does This Paper Address?
As Web3 loyalty programs and digital credentials move toward high-throughput, low-fee ledgers like Hedera or Polygon, a new challenge emerges: the "cost of attack" drops significantly. If it costs pennies to mint an identity and claim a reward, "farmers" can easily automate thousands of fraudulent claims.
The authors address the mechanism design problem of how an issuer can reward genuine participation without making "Sybil farming"—creating multiple identities to extract rewards—profitable. The core difficulty lies in the fact that measures to stop farmers (like long vesting periods or intrusive verification) often create friction that drives away legitimate users.
Methods
The authors develop a game-theoretic model where a "farmer" chooses the number of identities ($n$) to maximize their expected payoff. The model introduces three key defensive levers:
- Vesting ($T$): Rewards are not paid immediately but mature over time.
- Stochastic Verification ($a$): The issuer audits a random fraction of claims during the vesting period.
- Cluster Detection and Penalties ($F$): Crucially, the model assumes identities can be linked (via device fingerprints, funding paths, etc.). If one identity in a cluster is caught, the entire cluster is forfeited and a penalty is applied.
By using a continuous relaxation of the farmer's payoff function, the authors derive a "deterrence frontier"—the set of policies $(a, T)$ that make the optimal farming scale zero. They then solve the issuer's optimization problem: finding the mix of vesting and auditing that achieves deterrence at the lowest total cost (balancing audit expenses against the "friction cost" of delayed rewards for real users).
Datasets
Not discussed in the paper. This is a theoretical mechanism design paper; the authors use a calibrated numerical illustration rather than an empirical dataset to demonstrate their findings.
Key Insights and Ideas
- The Power of Linkage: In a standard account-level model, a farmer's payoff is linear—if one account is profitable, a million are. By modeling cluster-level detection, the authors turn the attack into a non-linear problem. As the farmer adds identities, their rewards grow linearly, but their probability of the entire cluster surviving detection drops geometrically. This creates a finite "optimal" attack size, which the issuer can then push to zero.
- Asymmetric Persistence: The model assumes genuine users are more "persistent" than farmers ($q_G > q_F$). This means vesting acts as a natural filter; farmers are more likely to abandon accounts or fail simple ongoing eligibility checks before the reward matures.
- Vesting Alone is Suboptimal: A major finding is that if the marginal cost of starting an audit program is low, relying solely on long vesting periods is never optimal. A small amount of active auditing significantly reduces the required vesting time, which helps retain genuine users who might otherwise be discouraged by long delays.
- Penalties vs. Auditing: The authors distinguish the roles of these two tools. A penalty ($F$) makes detection costly at any scale, while the audit rate ($a$) determines how quickly the risk of detection scales with the number of identities.
Key Results
- The Deterrence Condition: The paper derives a closed-form inequality for deterrence: $q_F^T R \le k_F + T L(a) F$. This compares the expected reward against the cost of creating identities plus the "hazard" of losing the penalty $F$ over the vesting window $T$.
- Numerical Optimization: In a calibrated example (Table 2), the authors show that an optimal mix of auditing and vesting can reduce the total "cost" of the mechanism (in terms of user friction and issuer spend) by roughly 50% compared to a vesting-only approach.
- Audit Intensity: The results show that as the cost of auditing ($d$) increases, the optimal strategy shifts toward longer vesting horizons ($T^*$) and lower audit rates ($a^*$). Conversely, if penalties ($F$) are high, the issuer can afford shorter vesting periods.
Limitations and Future Work
- Perfect Linkage Assumption: The model assumes that catching one Sybil identity always exposes the whole cluster. In reality, linkage is often partial or probabilistic. The authors acknowledge that weaker linkage would move the model back toward the account-level benchmark where deterrence is much harder.
- Risk Neutrality: The model assumes farmers are risk-neutral. Risk-averse farmers would be deterred more easily, while risk-seeking ones might require harsher penalties.
- False Positives: The baseline model assumes genuine claims never fail audits. The authors provide a brief extension for false positives, noting they significantly complicate the participation constraint for real users.
- Endogenous Rewards: The reward value $R$ is treated as a fixed constant. Future work could explore how the reward size itself should be tuned alongside the defense mechanisms.
Summary updated