homeblog aboutcontact

the fathom whitepaper:


The fathom protocol derives security from both it’s technical implementation and incentive structure. There are specific attack vectors that it mitigates against as well as some areas for further research.

Incentive Structure

While traditional credentials are meaningful because they are backed by reputable institutions, a fathom credential is meaningful because it is the result of many individuals having undergone a financial risk in the assessment game in order to create it. This section will lay out the decisions fathom users face when participating in that game, as well as the economic risks associated to them. Specifically, we show i) when assessors are likely to participate in an assessment in the first place and ii) why they can not collude with each other in order to shortcut the work associated with a truthful assessment.

Incentives for members to confirm or decline an assessment

This case is especially relevant for creating new concepts - as those will be initially empty and rely on members of its parent to participate in as assessors. A member of a concept who has received an offer to be an assessors will consider whether…

Incentives for assessors to grade truthfully

As truthfully assessing someone requires effort and the assessors payout is pegged to their alignment with each other, there is a motivation for them to collude, e.g. by agreeing ahead of time which score to commit. The creation of an adverserial environment between assessors is thus vital for the protocol to function as intended. Therefore, several mechanisms are put in place: First, assessors are paid out more if there are dissenting assessors ( figure 2). Consequently, any assessor taking part in a collusion of X assessors, must be afraid that they will be double crossed by a subgroup of more than \(X/2\) assessors. These are motivated to do so because they would be rewarded with part of the crossed assessors stake. Moreover, it is not possible for assessors to credibly prove to another assessor that they have actually committed to a collusion and logged in a previously agreed-upon score. In order to do so, the proving party would have to reveal their score and salt to the other assessors. Yet with this information, the other assessors could simply steal the assessors stake, which would eliminate the former assessor from the assessment and directly transfer the half of the revealed assessor’s stake to the revealer.

Attack Vectors

This section will outline some of the general classes of attacks against the protocol and a subjective view of their complexity, severity and to what degree they are considered to be mitigated.

Sybil Attacks

In a sybil attack, the attacker creates many false identities and then uses them to subvert the system, e.g. by controlling most of the identities in a concept, giving him control over who will be accepted and the ability to create assessments for himself in order to steal the stakes of other assessors.

To set up such an attack the attacker would, instead of being assessed by many assessors in one assessment, create multiple assessments with fewer assessors. This would be the same amount of work but result in four identities in the concept. Repeating the procedure, the attacker could count on some of his identities being called as assessors in which case the subsequent repetitions would become cheaper and less time-consuming until they have the majority in the concept or are called multiple times as assessor such that they can set up a 51% attack on individual assessments. In such a scenario, the attacker could control the outcome of the assessment and steal the stake of the other assessors.

Severity of attack: While a sybil attack does cost a fair amount of money to set up, the potential benefits are big enough to incentivize a try. As a compromised concept can potentially ‘poison’ its parent concepts as well and thus potentially effect the entire tree, we consider it to very severe.

Complexity: While a sybil attack is fairly complex, it can be effectuated by a single attacker, which is why it would be careless to assume that the degree of complexity will be a deterrent factor.

Degree of Protection: One possible mitigation that is not yet part of the protocol, will be to split the certificate and the right to be an assessor in two separate assessments. While this does not address the fundamental issue, it makes it easier for the sybil-protection measures to be integrated into the assessment process. For example, the assessment to become an assessor could ask the to-be-assessors for some piece of their own work or something that is new and can not be readily found on the internet as would be the case with the mere knowledge or skill required in the concept.

Simple Trolling

A troll, for arbitrary reasons, might try to poison the fathom network by creating a bunch of bogus assessments or concepts or by behaving irrationaly while being an assessor. In all cases, such behavior is expensive and ineffective, as his stakes are burned (when not following through with an assessment) or redistributed to others (when logging in bogus scores).

Bogus concepts will simply incur costs on the troll and be filtered out by assessors (see incentives). Creating bogus assessments as assessee will be even more costly (transaction costs and the fees for assessors). The worst effect a troll can have is to become an assessors and to prematurely end the assessment, if as a consequence of their behavior, its size is reduced below the minimum of five. In that case all other participants will be refunded, though.

Complexity: Behaving irrationaly is simple and so is attacking the system this way.

Severity: With no financial costs to other participants these kinds of attacks are not considered severe. An exception might be the creation of concepts, which if done by a well-resourced attacker, amounts to spamming the system.

Degree of protection: We consider simple trolling to be sufficiently discouraged because of the associated costs. If such behavior would be escalated into a spam-attack of greater proportions, the degree of protection will depend on the users or the fathom frontends ability to filter concepts and assessments by meaningful criteria.

P + epsilon attack

In a P + epsilon attack, the attacker circumvents the incentivization by creating a mechanism that others can trust in because it gives them a credible guarantee about the attacker’s behavior. While this would have been difficult in a pre-blockchain era, smart contracts are nearly ideally suited to implement such mechanisms.

The attack works like this: In a schelling-point game, the assessors are being paid out the same amount \(P\). regardless of the result (option A, B, C or any other...). The attacker, let’s say wanting to push for a certain option A, will credibly guarantee anyone voting for A that he will be paid P+\(\epsilon\), if they vote A and the majority doesn’t. Assuming a system that is not dominated by altruistic actors, voting A is now the game-theoretically best option (guaranteed maximal payout). Therefore, the majority will vote A and the attacker will have taken over the mechanism - at zero cost.

Although there exist some protection mechanisms that can increase the attackers risk (size of the needed bribe) and some counter-coordination mechanisms that come close to defeating such an attack, there is currently no guaranteed countermeasure.

Complexity: As the crucial element of this attack is the mechanism by which the attacker commits to his intention to paying out in case the bribed voter is not in the majority, the complexity is proportional to the difficulty of construing such a mechanism. In the case of fathom, the difficulty to reconstruct the relevant information (did an assessor really vote for the desired option A?). Currently, this is rather simple, so setting up this attack would not be very complex.

Severity: As this attack can disrupt the system at potentially zero-cost, we consider it to be very severe.

Degree of Protection: As of right now, the protocol is not protected against such measurs. Future versions of it could implement some more complicated schemes in order to keep the scores of individual assessors secret and make it harder to retrieve the individual assessors’ scores.