Sybil resistance

The attack, and the result that constrains every defense

John Douceur named the attack at Microsoft Research in 2002 and, in the same paper, drew the boundary around every answer to it. Without a logically centralized identity authority, sybil attacks remain possible except under extreme and unrealistic assumptions about resource parity and coordination between entities.¹

Read that as a design constraint rather than a reason to give up. A fully decentralized system cannot make sybil attacks impossible, so the goal is to make them uneconomic for the specific thing being protected. The value at stake sets the budget. A governance vote controlling a treasury deserves far more resistance than a testnet leaderboard, and spending equally on both is how teams end up with a distribution nobody qualifies for and a vote anyone can buy.

Four approaches, four currencies

Cost makes each identity buy something scarce. Proof of work is the pure form and the currency is energy; a per-address fee, a minimum balance or a gas-heavy qualifying action is the cheap form. Resistance scales with the price of the resource, and so does the burden on honest participants.

Stake weights by capital at risk instead of by head. Proof of stake is sybil resistant by construction, since splitting one stake across a thousand addresses buys nothing. It converts a sybil problem into a concentration problem, which is a trade rather than a solution.

Proof of personhood binds one credential to one human, verified by an authority, a document or a device. It offers the most resistance per unit of friction and carries the largest privacy and access cost. Social graph analysis infers uniqueness from who vouches for whom, or from behavioral patterns. No hardware, no document, and it degrades as soon as the graph itself becomes worth farming.

Proof of personhood and the dependency it creates

World ID publishes a proof of human credential with an SDK and on-chain verification, which is the shape most teams evaluate in this category.² The mechanism is straightforward. The consequence is the part worth thinking about.

Adopting an external personhood provider imports a third party into your distribution. Eligibility now depends on that provider's country coverage, uptime and policy, in every jurisdiction your users live in. That is a dependency of the same class as an oracle and it belongs in the risk register next to one. For a global airdrop it also decides who cannot participate, which is a distribution decision made by somebody else's onboarding footprint rather than by your allocation table.

Airdrop sybil detection, where most readers arrive

Detection is the after-the-fact version: clustering rather than identification. The Arbitrum Foundation published its method: on-chain data used to identify related addresses owned by the same user, with entity addresses such as bridges, exchanges and smart contracts removed using data from Nansen, Hop and Offchain Labs, then the resulting graphs partitioned into strongly and weakly connected subgraphs, with large subgraphs broken down using the Louvain community detection algorithm. The write-up names specific clusters with their eligible address counts, including Cluster 319 with 110 and Cluster 2554 with 121.³

Two things follow. Clustering produces judgments, not facts, so an appeals path is part of the design and not a support queue invented on the day. And every criterion you publish is a specification for the next farmer, which argues for publishing the method after the snapshot, not before.

What we tell teams before a distribution

Start by naming what you are protecting and what one stolen unit costs you. Size the resistance to that number, not to what other projects did. Then take the snapshot at a block nobody could have known in advance, since a published date converts your eligibility rule into a shopping list.

Across the 80+ projects we have advised, the distributions that hold up are the ones where the qualifying behavior costs a farmer roughly what it costs a genuine user. That is a design property, not an enforcement one. Sybil resistance bolted on after the snapshot is a cleanup operation with a public relations problem attached, and it ends with a team defending a spreadsheet to people who believe they were wrongly excluded.