Security.rst 9.01 KB
Newer Older
Jared Pereira's avatar
Jared Pereira committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
.. _sec:security:


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

.. _sec:incentive:

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.

.. _sec:acceptAss:

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…

-  They feel competent enough in their abilities to come to the same
   conclusion as a group of other, randomly selected assessors that are
   confident in *their* abilities.

-  The concept in question is well enough defined so that assessors with
   similar impressions of the assessee’s skill will be able to translate
   these into similar scores.

.. _sec:notCollude:

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 (see `Figure
1 <./2-Protocol.html#sec:payout>`__) 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 :math:`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 <./2-Protocol.html#sec:stealReveal>`__ 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

.. _sec:attackVectors:

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.

.. _sec:sybil:

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

**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.

.. _sec:trolling:

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 <#sec:acceptAss>`__). 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.

.. _sec:blackmail:

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 :math:`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+\ :math:`\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.