M ECHANISM D ESIGN FOR M ULTI -PARTY M ACHINE

L EARNING
Mengjing Chen∗ 1 , Yang Liu† 2 , Weiran Shen‡ 3 , Yiheng Shen§ 1 , Pingzhong Tang¶ 1 , and Qiang
Yangk 4
1
Tsinghua University
2
WeBank
3
arXiv:2001.08996v3 [cs.MA] 9 Aug 2020

Carnegie Mellon University

4
Hong Kong University of Science and Technology

A BSTRACT

In a multi-party machine learning system, different parties cooperate on optimizing

towards better models by sharing data in a privacy-preserving way. A major chal-
lenge in learning is the incentive issue. For example, if there is competition among
the parties, one may strategically hide his data to prevent other parties from getting
better models.
In this paper, we study the problem through the lens of mechanism design and
incorporate the features of multi-party learning in our setting. First, each agent’s
valuation has externalities that depend on others’ types and actions. Second, each
agent can only misreport a type lower than his true type, but not the other way
round. We call this setting interdependent value with type-dependent action spaces.
We provide the optimal truthful mechanism in the quasi-monotone utility setting.
We also provide necessary and sufficient conditions for truthful mechanisms in the
most general case. Finally, we show the existence of such mechanisms is highly
affected by the market growth rate and provide empirical analysis.

1 Introduction
In multi-party machine learning, a group of parties cooperates on optimizing towards better mod-
els. This concept has attracted much attention recently [13, 24, 25]. The advantage of this approach
is that, it can make use of the distributed datasets and computational power to learn a powerful
model that anyone in the group cannot achieve alone.
To make multi-party machine learning practical, a large body of works focus on preserving data
privacy in the learning process [1, 26, 24]. However, the incentive issues in the multi-party learning
have largely been ignored in most previous studies, which results in a significant reduction in the
∗
[email protected]
†
[email protected]
‡
[email protected]
§
[email protected]
¶
[email protected]
k
[email protected]
effectiveness when putting their techniques into practice. Previous works usually let all the parties
share the same global model with the best quality regardless of their contributions. This allocation
works well when there are no conflicts of interest among the parties. For example, an app developer
wants to use the users’ usage data to improve the user experience. All users are happy to contribute
data since they can all benefit from such improvements [15].
When the parties are competing with one another, they may be unwilling to participate in the
learning process since their competitors can also benefit from their contributions. Consider the
case where companies from the same industry are trying to adopt federated learning to level up
the industry’s service qualities. Improving other companies’ services can possibly harm their own
market share, especially when there are several monopolists that own most of the data.
Such a cooperative and competitive relation poses an interesting challenge that prevents the multi-
party learning approach from being applied to a wider range of environments. In this paper, we
view this problem from the multi-agent system perspective, and address the incentive issues men-
tioned above with the mechanism design theory.
Our setting is a variant of the so-called interdependent value setting [19]. A key difference be-
tween our setting and the standard interdependent value setting is that each agent cannot “make
up” a dataset that is of higher quality than his actual one. Thus the reported type of an agent
is capped by his true type. We call our setting interdependent value with type-dependent action
spaces. The setting that agents can never over-report is common in practice. One straightforward
example is that the sports competitions where athletes can show lower performance than their ac-
tual abilities but not over-perform. The restriction on the action space poses more constraints on
agents’ behaviors, and allows more flexibility in the design space.
We first formulate the problem mathematically, and then apply techniques from the mechanism
design theory to analyze it. Our model is more general than the standard mechanism design frame-
work, and is also able to describe other similar problems involving both cooperation and competi-
tion.
We make the following contributions in this paper:

• We model and formulate the mechanism design problem in multi-party machine learning,
and identify the differences between our setting and the other mechanism design settings.

• For the quasi-monotone externalities setting, we provide the revenue-optimal and truthful
mechanism. For the general valuation functions, we provide both the necessary and the
sufficient conditions for all truthful and individually rational mechanisms.

• We analyze the influence of the market size on mechanisms. When the market grows
slowly, there may not exist a mechanism that achieves all the desirable properties we focus
on.

1.1 Related Works

A large body of literature studies mechanisms with interdependent values [19], where agents’ val-
uations depend on the types of all agents and the intrinsic qualities of the allocated object. [23]

2
extend Myerson’s auction to specific interdependent value settings and characterize truthful and ra-
tional mechanisms. They consider a bayesian setting while we do not know any prior information.
[5] propose a variant of the VCG auction with reserve prices that can achieve high revenues. They
consider value functions that are single-crossing and concave while we consider environments with
more general value functions. [18] give a two-stage Groves mechanism that guarantees truthful-
ness and efficiency. However, they require agents to report their types and valuations before the
final monetary transfer are made while in our model, agents can only report their types.
In our setting, agents have restricted action spaces, i.e., they can never report types exceeding
their actual types. There is a series of works that focus on mechanism design with a restricted
action space [3, 4, 2]. The discrete-bid ascending auctions [8, 6, 2] specify that all bidders’ action
spaces are the same bid level set. Several works restrict the number of actions, such as bounded
communications [4]. Previous works focus on mechanisms with independent values and discrete
restricted action spaces, while we study the interdependent values and continuous restricted action
spaces setting.
The learned model can be copied and distributed to as many agents as possible, so the supply is
unlimited. A line of literature focuses on selling items in unlimited supply such as digital goods
[12, 11, 10]. However, the seller sells the same item to buyers while in our setting we can allocate
models with different qualities to different agents.
[22] study the optimal strategies of agents for collaborative machine learning problems. Both their
work and ours capture the cooperation and competition among the agents, but they only consider
the case where agents reveal their total datasets to participate while agents can choose to contribute
only a fraction in our setting. [14] study the incentive design problem for federated learning,
but all their results are about a non-competitive environment, which may not hold in real-world
applications.
Another closely related topic is the incentives in machine learning problems, such as strategyproof
classification [17, 16], incentive compatible linear regression [7, 9, 21]. The main objective of
these mechanisms is to get an unbiased learning model with strategic agents and all agents receive
the results from the same model. We also consider the collaborative learning setting, but different
models are distributed to different agents.

2 Preliminaries

In this section, we introduce the general concepts of mechanism design and formulate the multi-
party machine learning as a mechanism design problem. A multi-party learning consists of a
central platform and several parties (called agents hereafter). The agents serve their customers
with their models trained using their private data. Each agent can choose whether to enter the
platform. If an agent does not participate, then he trains his model with only his own data. The
platform requires all the participating agents to contribute their data in a privacy-preserving way
and trains a model for each participant using a (weighted) combination of all the contributions.
Then the platform returns the trained models to the agents.
We assume that all agents use the same model structure. Therefore, each participating agent may
be able to train a better model by making use of his private data and the model allocated to him.

3
One important problem in this process is the incentive issue. For example, if the participants have
conflicts of interest, then they may only want to make use of others’ contributions but are not
willing to contribute with all their own data. To align their incentives, we allow the platform to
charge the participants according to some predefined rules.
Our goal is to design allocation and payment rules that encourage all agents to join the multi-party
learning as well as to contribute all their data.

2.1 Valid Data Size (Type)

Suppose there are n agents, denoted by N = (1, 2, . . . , n), and each of them has a private dataset
Di where Di ∩ Dj = ∅, ∀i 6= j. For ease of presentation, we assume that a model is fully
characterized by its quality Q (e.g., the prediction accuracy), and the quality only depends on the
data used to train it. We have the following observation:
Observation 1. If the agents could fake a dataset with a higher quality, any truthful mechanism
would make agents gain equal final utility.
Suppose that two agents have different true datasets D1 and D2 . We assume that all other agents
truthfully report datasets D−i . If truthfully reporting dataset D1 and D2 finally leads to different
utility, w.l.o.g, we let u(D1 , D−i ) < u(D2 , D−i ), then if an agent has true dataset D1 , he would
report D2 and use the dataset allocated by the platform in the market. All his behavior is the same
as that of an agent with real dataset D2 . Thus if a mechanism is truthful, any reported dataset
would lead to the same final utility and thus it is pointless to discuss the problem. Hence, we make
the assumption that the mechanism is able to identify the quality of any dataset. All agents can
only report a dataset with a lower quality.
For simplicity, we measure the contribution of a dataset to a trained model by its valid data size.
Thus we have the following assumption:
Assumption 1. The model quality Q is bounded and monotone increasing with respect to the valid
data size s ≥ 0 of the training data:

1. Q(0) = 0 and Q(s) ≤ 1, ∀s;

2. Q(s0 ) > Q(s), ∀s0 > s.

The valid data size of every contributor’s data is validated by the platform in a secure protocol
(which we propose in the full version ). Let ti ∈ R+ be the valid data size of agent i ’s private
dataset Di . We call ti the agent’s type. The agent can only falsify his type by using a dataset
of lower quality (for example, using a subset of Di , or adding fake data), which decreases the
contribution to the trained model as well as the size of valid data. As a result, the agent with type
ti cannot contribute to the platform with a dataset higher than his type:
Assumption 2. Each agent i can only report a type lower than his true type ti , i.e., the action
space of agent i is [0, ti ].

2.2 Mechanism

Let t = (t1 , t2 , . . . ,tn ) and t−i = (t1 , . . . , ti−1 , ti+1 , . . . , tn ) be the type profile of all agents and
all agents without i, respectively. Given the reported types of agents, a mechanism specifies a

4
numerical allocation and payment for each agent, where the allocation is a model in the multi-
party learning. Formally, we have:
Definition 1 (Mechanism). A mechanism M = (x, p) is a tuple, where
x = (x1 , x2 , · · · , xn ), where xi : Rn+ 7→ R takes the agents’ reported types as input and outputs the
model quality for agent i;
p = (p1 , p2 , · · · , pn ), where pi : Rn+ 7→ R takes the agents’ reported types as input and specifies
how much agent i should pay.
In a competitive environment, a strategic agent may hide some of data and does not use the model
he receives from the platform. Thus the final model quality depends on both the allocation and his
actual type. We use valuation function vi (x(t0 ), t) to measure the profit of agent i.
Definition 2 (Valuation). We consider valuation functions vi (x(t0 ), t) that depend not only on the
allocation outcome x(t0 ) where t0 is the reported type profile, but also on the actual type profile t.
We assume the model agent i uses to serve customers is:

qi = max{xi (t0 ), Q(ti )},

where Q(ti ) is the model trained with his own data. The valuation of agent i depends on the
final model qualities of all agents due to their competition. Hence vi can also be expressed as
vi (q1 , . . . , qn ).
We make the following assumption on agent i’s valuation:
Assumption 3. Agent i’s valuation is monotone increasing with respect to true type ti when the
outcome x is fixed.

vi (x, ti , t−i ) ≥ vi (x, t̂i , t−i ), ∀x, ∀ti ≥ t̂i , ∀t−i , ∀i.

This is because possessing more valid data will not lower one’s valuation. Otherwise, an agent is
always able to discard part of his dataset to make his true type t0i . Suppose that each agent i’s utility
ui (t, t0 ) has the form:

ui (t, t0 ) = vi (x(t0 ), t) − pi (t0 ),

where t and t0 are true types and reported types of all agents respectively. As mentioned above,
an agent may lie about his type in order to benefit from the mechanism. The mechanism should
incentivize truthful reports to keep agents from lying.
Definition 3 (Incentive Compatibility (IC)). A mechanism is said to be incentive compatible, or
truthful, if reporting truthfully is always the best response for each agent when the other agents
report truthfully:

ui (x(ti , t−i ), t) ≥ ui (xi (t0i , t−i ), t), ∀ti ≥ t0i , ∀t−i , ∀i.

For ease of presentation, we say agent i reports ∅ if he chooses not to participate (so we have
xi (∅, t−i ) = 0 and pi (∅, t−i ) = 0). To encourage the agents to participate in the mechanism, the
following property should be satisfied:

5
Definition 4 (Individual Rationality (IR)). A mechanism is said to be individually rational, if no
agent loses by participation when the other agents report truthfully:
ui (x(ti , t−i ), t) ≥ ui (x(∅, t−i ), t), ∀ti , t−i , ∀i.

The revenue and welfare of a mechanism are defined to be all the payments collected from the
agents and all the valuations of the agents.
Definition 5. The revenue and welfare of a mechanism (x, p) are:
n
X
R EV(x, p) = pi (t0 ),
i=1
n
X
W EL(x, p) = vi (x, t).
i=1

We say that a mechanism is efficient if

(x, p) = arg max W EL(x, p),
(x,p)

A mechanism is weakly budget-balance if it never loses money.

Definition 6 (Weak Budget Balance). A mechanism is weakly budget-balance if:
R EV(x, p) ≥ 0, ∀t.
Definition 7 (Desirable Mechanism). We say a mechanism is desirable if it is IC, IR, efficient and
weakly budget-balance.

2.3 Comparison with the Standard Interdependent Value Setting

Although each agent’s valuation depend on both the outcome of the mechanism and all agent’s true
types, our interdependent value with type-dependent action spaces setting, however, fundamentally
different from standard interdependent value settings:

• In our setting, the type of each agent is the “quality” of his dataset, thus each agent cannot
report a higher type than his true type. While in the standard interdependent value setting,
an agent can possibly report any type.
• In our setting, the agents do not have the “exit choice” (not participating in the mechanism
and getting 0 utility) as they do in the standard setting. This is due to the motivation of this
paper: companies from the same industry are trying to improve their service quality, and
they are always in the game regardless of their choices. A non-participating company may
even have a negative utility if all other companies improved their services.
• To capture the cooperation among the agents, the item being sold, i.e., the learned model,
also depends on all agents types. The best model learned by the multi-party learning plat-
form will have high quality if all agents contribute high-quality datasets. However, the
objects for allocation are usually fixed in standard mechanisms instead.

6
3 Quasi-monotone Externality Setting
In the interdependent value with type-dependent action spaces setting, each agent’s utility may
also depend on the models that other agents actually use. Such externalities lead to interesting and
complicated interactions between the agents. For example, by contributing more data, one may
improve the others’ model quality, and end up harming his own market share.
In this section, we study the setting where agents have quasi-monotone externalities.
Definition 8 (Quasi-Monotone Valuation). Let qi be the final selected model quality of the agent
and q−i be the profile of model qualities of all the agents except i. A valuation function is quasi-
monotone if it is in the form:
vi (qi , q−i ) = Fi (qi ) + θi (q−i ),
where Fi is a monotone function and θi is an arbitrary function.
Example 1. Let’s consider a special quasi-monotone Pvaluation: the linear externality setting,
where the valuation for each agent is defined as vi = j αij qj with qj being the model that agent
j uses. The externality coefficient αij means the influence of agent j to agent i and captures either
the competitive or cooperative relations among agents. If the increase of agent j’s model quality
imposes a negative effect on agent i’s utility (e.g. major opponents in the market), αij would be
negative. Also aij could be positive if agent i and agent j are collaborators. Additionally, αii
should always be positive, naturally.

P the linear externality setting, the efficient allocation is straightforward. For each agent i, if
In
j αji ≥ 0, the platform gives agent P
i the training model with best possible quality. Otherwise,
agent i are not allocated any model if j αji < 0.

We introduce a payment function called maximal exploitation payment, and show that the mecha-
nism with efficient allocation and the maximal exploitation payment guarantees individually ratio-
nality, truthfulness, efficiency and revenue optimum.
Definition 9 (Maximal Exploitation Payment (MEP)). For a given allocation function x, suppose
the agent i reports a type t0i and the other agents report t0−i , the maximal exploitation payment is
to charge i
pi (t0i , t0−i ) = vi (x(t0i , t0−i ), t0i , t0−i ) − vi (x(∅, t0−i ), t0i , t0−i ).
Theorem 1. Under the quasi-monotone valuation setting, any mechanism with MEP is the mech-
anism with the maximal revenue among all IR mechanisms, and it is IC.

Proof. Intuitively, the MEP rule charges agent i the profit he gets from an model that the mech-
anism allocates to him. If the mechanism charges higher than the MEP, an agent would have
negative utility after taking part in. The IR constraint would then be violated. So it’s easy to see
that the MEP is the maximal payment among all IR mechanisms.
Then we prove that this payment rule also guarantees the IC condition. It suffices to show that if an
agent hides some data, no matter which model he chooses to use, he would never get more utility
than that of truthful reporting. We suppose that agent i’s type is t0i and he untruthfully reports t0i .
Suppose that the agent i truthfully reports the type t0i = ti , since the payment function is defined to
charge this agent until he reaches the valuation when he does not take part in the mechanism, the

7
utility of this honest agent would be

u0i (t0 ) = Fi (Q(ti )) + θi (q−i (∅, t0−i )).

If the agent does not report truthfully, we suppose that the agent reports t0i where t0i ≤ ti . According
to the MEP, the payment function for agent i would be

pi (t0i , t0−i ) = Fi (qi (t0i , t0−i )) + θi (q−i (t0i , t0−i )) − Fi (Q(t0i )) − θi (q−i (∅, t0−i )).

It can be seen that the mechanism would never give an agent a worse model than the model trained
by its reported data, otherwise the agents would surely select their private data to train models.
Hence it is without loss of generality to assume that the allocation xi (t0i , t0−i ) ≥ Q(t0i ), ∀t0i , t0−i , ∀i.
Thus we have q−i (t0i , t0−i ) = x−i (t0i , t0−i ). We discuss the utility of agent i by two cases of choosing
models.
Case 1: the agent chooses the allocation xi . Since agent i selects the allocated model, we have
qi = xi (t0i , t0−i ). Then the utility of agent i would be

u1i =vi (t0i , t0−i ) − pi (t0i , t0−i )

=Fi (xi (t0i , t0−i )) + θi (x−i (t0i , t0−i )) − pi (t0i , t0−i )
=Fi (xi (t0i , t0−i )) + θi (x−i (t0i , t0−i )) + Fi (Q(t0i ))
+ θi (x−i (∅, t0−i )) − Fi (xi (t0i , t0−i )) − θi (x−i (t0i , t0−i ))
=Fi (Q(t0i )) + θi (x−i (∅, t0−i )).

Because both Fi and Q are monotone increasing functions and ti ≥ t0i , we have u1i ≤ Fi (Q(ti )) +
θi (x−i (∅, t0−i )) = u0i .
Case2: the agent chooses Q(ti ). Since agent i selects the model trained by his private data, we
have qi = Q(ti ). The final utility of agent i would be

u2i =vi (t0i , t0−i ) − pi (t0i , t0−i )

=Fi (Q(ti )) + θi (x−i (t0i , t0−i )) − pi (t0i , t0−i )
=Fi (Q(ti )) + θi (x−i (t0i , t0−i )) + Fi (Q(t0i ))
+ θi (x−i (∅, t0−i )) − Fi (xi (t0i , t0−i )) − θi (x−i (t0i , t0−i ))
=Fi (Q(ti )) + Fi (Q(t0i )) + θi (x−i (∅, t0−i )) − Fi (xi (t0i , t0−i )).

Subtract the original utility from the both sides, then we have

u2i − u0i =Fi (Q(ti )) + Fi (Q(t0i )) + θi (x−i (∅, t0−i ))

− Fi (xi (t0i , t0−i )) − Fi (Q(ti )) − θi (x−i (∅, t0−i ))
=Fi (Q(t0i )) − Fi (xi (t0i , t0−i )).

Because xi (t0i , t0−i ) ≥ Q(t0i ), ∀t0i , t0−i , ∀i and because Fi is a monotonically increasing function, we
can get u2i − u0i ≤ 0. Therefore max{u1i , u2i } ≤ u0i , lying would not bring more benefits to any
agent, and the mechanism is IC.

8
Corollary 1. Any efficient allocation mechanism with MEP under the linear externality setting
with all the linear coefficients αji ≥ 0 should be IR, IC, weakly budget-balance and efficient.

Proof. In Theorem 1 we know that the MEP mechanism is IR and IC. Since the linear coefficients
are all positive and the externality setting is linear, any efficient mechanism would allocate the best
model to all the agents. Since each agent gets a model with no less quality than his reported one
and the payment is equal to the value difference between the case an agent truthfully report and
the case he exit the mechanism. The agent’s value is always larger than the value when he exits the
mechanism. Then the payment is always positive and the mechanism should satisfy all of the four
properties.

In the standard mechanism design setting, the Myerson-Satterthwaite Theorem [20] is a well-
known classic result, which says that no mechanism is simultaneously IC, IR, efficient and weakly
budget-balance. The above Corollary 1 shows that in our setting, the Myerson-Satterthwaite
Theorem fails to hold.

4 General Externality Setting

In this section, we consider the general externality setting where the valuation of agent i can be
any function of allocation outcome and types of all agents. The limitation on the reporting space
and the type-imposed value functions make the IC and IR mechanisms hard to characterize. It is
possible that given a allocation rule, there exist several mechanisms with different payments that
satisfy both IC and IR constraints. To understand what makes a mechanism IC and IR, we analyze
some properties of truthful mechanisms in this section.
For ease of presentation, we assume that the functions v(·), x(·) and p(·) are differentiable.
Theorem 2 (Necessary Condition). If a mechanism (x, p) is both IR and IC, for all possible valu-
ation functions satisfying Assumption 3, then the payment function satisfies ∀ti ≥ t0i , ∀ti , ∀t−i , ∀i,

pi (0, t−i ) ≤ vi (x(0, t−i ), 0, t−i ) − vi (x(∅, t−i ), 0, t−i ), (1)

Z ti
0 ∂vi (x(s0 , t−i ), s, t−i )
pi (ti , t−i ) − pi (ti , t−i ) ≤ 0
ds0 , (2)
0
ti ∂s s=s0

where we view vi (x(t0i , t−i ), ti , t−i ) as a function of ti , t0i and t−i for simplicity and
n
∂vi (x(t0i , t−i ), ti , t−i ) X ∂vi (x(t0i , t−i ), ti , t−i ) ∂xj (t0i , t−i )
= .
∂t0i i=1
∂xj (t0i , t−i ) ∂t0i

Proof. We first prove that Equation (1) holds. Observe that

ui (x(ti , t0−i ), ti , t−i ) − ui (x(t0i , t0−i ), t0i , t−i )

=[vi (x(ti , t0−i ), ti , t−i ) − pi (ti , t0−i )] − [vi (x(t0i , t0−i ), t0i , t−i ) − pi (t0i , t0−i )]
≥[vi (x(ti , t0−i ), ti , t−i ) − pi (ti , t0−i )] − [vi (x(t0i , t0−i ), ti , t−i ) − pi (t0i , t0−i )]
=ui (x(ti , t0−i ), ti , t−i ) − ui (x(t0i , t0−i ), ti , t−i )
≥0, (3)

9
where the first inequality is because of Assumption 3, and the last inequality is because of the
DSIC property.
Let t0i = 0 in Equation (3). We have
ui (x(ti , t0−i ), ti , t−i ) ≥ ui (x(0, t0−i ), 0, t−i ).
The IR property further requires that ui (x(0, t0−i ), 0, t−i ) ≥ ui (x(∅, t0−i ), 0, t−i ), which Equation
(1) follows.
To show Equation (2) must hold, we rewrite Equation (3):
pi (ti , t0−i ) − pi (t0i , t0−i ) ≤vi (x(ti , t0−i ), ti , t−i ) − vi (x(t0i , t0−i ), t0i , t−i )
dvi (x(s0 , t0−i ), s(s0 ), t−i ) 0
Z ti
= ds . (4)
t0i ds0

Fixing t−i and t0−i , the total derivative of the valuation function vi (x(s0 , t0−i ), s, t−i ) is:
∂vi (x(s0 , t0−i ), s, t−i ) 0 ∂vi (x(s0 , t0−i ), s, t−i )
dvi (x(s0 , t0−i ), s, t−i ) = ds + ds.
∂s0 ∂s
View s as a function of s0 and let s(s0 ) = s0 :
dvi (x(s0 , t0−i ), s(s0 ), t−i ) ∂vi (x(s0 , t0−i ), s, t−i ) ∂vi (x(s0 , t0−i ), s(s0 ), t−i ) ds(s0 )
= + .
ds0 ∂s0 s=s0 ∂s(s0 ) ds0
Plug into Equation (4), and we obtain:
∂vi (x(s0 , t0−i ), s, t−i ) ∂vi (x(s0 , t0−i ), s(s0 ), t−i ) 0
Z ti Z ti
0 0 0
pi (ti , t−i ) − pi (ti , t−i ) ≤ + ds .
t0i ∂s0 s=s0 t0i ∂s(s0 )
Since the above inequality holds for any valuation function with vi (x, ti , t−i ) ≥
vi (x, t0i , t−i ), ∀x, ∀t−i , ∀ti ≥ t0i , we have:
∂vi (x(s0 , t0−i ), s, t−i )
Z ti
0 0 0
pi (ti , t−i ) − pi (ti , t−i ) ≤ ds0 .
t0i ∂s0 s=s 0

Theorem 2 describes what the payment p is like in all IC and IR mechanisms. In fact, the conditions
in Theorem 2 are also crucial in making a mechanism truthful. However, to ensure IC and IR
property, we still need to restrict the allocation function x.
Theorem 3 (Sufficient Condition). A mechanism (x, p) satisfies both IR and IC, for all possible
valuation functions satisfying Assumption 3, if for each agent i, for all ti ≥ t0i , and all t−i , Equa-
tions (1) and the following two hold
∂vi (x(t0i , t−i ), ti , t−i )
t0i = arg min (5)
ti :ti >t0i ∂t0i
ti ti
∂vi (x(s0 , t−i ), s, t−i )
Z Z
∂vi (x(∅, t−i ), s, t−i )
pi (ti , t−i ) − pi (t0i , t−i ) ≤ 0
ds − ds. (6)
t0i ∂s0 s=s0 t0i ∂s

10
 ∂vi (x(t0i ,t0−i ),ti ,t−i )
Proof. Equation (5) indicates that the function ∂t0i
is minimized at t0i :

∂vi (x(t0i , t0−i ), s, t−i ) ∂vi (x(t0i , t0−i ), ti , t−i )

≤ . (7)
∂t0i s=t0i ∂t0i

Therefore, we have

ui (x(ti , t0−i ), ti , t−i ) − ui (x(t0i , t0−i ), ti , t−i )

∂vi (x(s0 , t0−i ), ti , t−i ) 0
Z ti
= 0
ds − pi (ti , t0−i ) + pi (t0i , t0−i )
t 0 ∂s
Z iti
∂vi (x(s0 , t0−i ), s, t−i )
≥ 0
ds0
t0i ∂s s=s0
− pi (ti , t0−i ) + pi (t0i , t0−i )
∂vi (x(∅, t0−i ), s, t−i )
Z ti
≥ ds, (8)
t0i ∂s

where the two inequalities are due to Equation (7) and (6), respectively. Since
0 0 ∂vi (x(∅,t0−i ),s,t−i )
vi (x, ti , t−i )≥vi (x, ti , t−i ), ∀x,∀t−i ,∀ti ≥ ti indicates ∂s
≥ 0, the above inequality
shows that the mechanism guarantees the DSIC property.
To prove that the mechanism is IR, we first observe that

[ui (x(ti , t0−i ), ti , t−i ) − vi (x(∅, t0−i ), ti , t−i )] − [ui (x(t0i , t0−i ), t0i , t−i ) − vi (∅, x(t−i ), t0i , t−i )]
∂vi (x(∅, t0−i ), s, t−i )
Z ti
0 0 0 0
=ui (x(ti , t−i ), ti , t−i ) − ui (x(ti , t−i ), ti , t−i ) − ds
t0i ∂s
∂vi (x(∅, t0−i ), s, t−i )
Z ti
0 0 0
≥ui (x(ti , t−i ), ti , t−i ) − ui (x(ti , t−i ), ti , t−i ) − ds
t0i ∂s
≥0,

where the two inequalities are Assumption 3 and Equation (8). Letting t0i = 0 using Equation (2),
we get:

ui (x(ti , t0−i ), ti , t−i ) − vi (x(∅, t0−i ), ti , t−i )

≥ui (x(0, t0−i ), 0, t−i ) − vi (x(∅, t0−i ), 0, t−i )
=vi (x(0, t0−i ), 0, t−i ) − pi (0, t0−i ) − vi (x(∅, t0−i ), 0, t−i )
≥0.

5 Market Growth Rate

In this section, we will analyze a factor, the market growth rate, for the existence of the desirable
mechanism. Expanding the market size would reduce competition among the agents, meaning that

11
the damage to an agent’s existing market caused by joining the mechanism is more likely to be
covered by the market growth. Thus our intuition is that if the market grows quickly, a desirable
mechanism is more likely to exist.
As mentioned above, each agent’s valuation is the profit made from the market, so formally we
define the market size to be the sum of the valuations of all the agents. Let M (q) be the agents’
total valuations where q = (q1 , q2 , . . . , qn ) is the set of actual model qualities they use. We have:
n
X
M (q) = vi (x, t).
i=1

In general, the multi-party learning process improves all agents’ models. So we do not consider
the case where the market shrinks due to the agents’ participation, and assume that the market is
growing.
Assumption 4 (Growing Market). q  q 0 implies M (q) ≥ M (q 0 ).

A special case of the growing market is the non-competitive market where agent’s values are not
affected by others’ model qualities, formally:
∂vi (q)
Definition 10 (Non-competitive Market). A market is non-com-petitive if and only if ≥
∂qj
0, ∀i, j.
Theorem 4. In a non-competitive market, there always exists a desirable mechanism, that gives
the best possible model to each agent and charges nothing.

Proof. Suppose that the platform uses the mechanism mentioned in the theorem. Then for each
agent, contributing with more data increases all participants’ model qualities. By definition, in a
non-competitive market, improving others’ models does not decrease one’s profit. Therefore, the
optimal strategy for each participant is to contribute with all his valid data, making the mecha-
nism truthful. Also because of the definition, entering the platform always weakly increases one’s
model quality. Thus the mechanism is IR. With the IC and IR properties, it is easy to see that the
mechanism is also efficient and weakly budget-balance.

However, when the competition exists in a growing market, it is not easy to determine whether
a desirable mechanism exists. Since when the market is growing, the efficient mechanism both
redistributes existing markets and enlarges the market size by giving the best learned model. We
will give the empirical analysis of the influence of competition and the market growth rate on the
existence of desirable mechanisms.

6 Experiments
We design experiments to demonstrate the performance of our mechanism for practical use. We
first show the mechanism with maximal exploitation payments (MEP) can guarantee good quality
of trained models and high revenues under the linear externality cases. Then we conduct simula-
tions to show the relation of the market growth of competitive market to the existence of desirable
mechanisms.

12
 1.0
avg_wel
avg_rev
Welfare/Revenue

300

Model quality
200 0.5
100

0 0.0 1
0 50 100 150 5 10 15 20
The number of agents The number of agents
(a) Welfare & revenue (b) Best quality of trained model

Figure 1: Performance of MEP under different numbers of agents

6.1 The MEP Mechanism

We consider the valuation with linear externalities setting where αij ’s (defined in Example ??) are
generated uniformly in [−1, 1]. Each agent’s type is drawn uniformly from [0, 1] independently
1−e−t
and the Q(t) is 1+e −t . The performance of a mechanism is measured by the platform’s revenue and

its best quality of trained model under the mechanism. All the values of each instance are averaged
over 50 samples. We both show the performance changes as the number of agents increases and as
the agents’ type changes.
When the number of agents becomes larger, the platform can obtain more revenues and train better
models (see Figure 1). Particularly, the model quality is close to be optimal when the number of
agents over 12. An interesting phenomenon is that the revenue may surpass the social welfare.
This is because the average external effect of other agents on one agent i tends to be negative when
agent i does not join in the mechanism, thus the second term in the MEP payment is averagely
negative and revenue is larger than the welfare.
To see the influence of type on performance, we fix one agent’s type to be 1 and set the other
agent’s type from 0 to 2. It can be seen in Figure 2 that the welfare and opponent agent’s utility
(uti_2) increase as the opponent’s type increases but the platform’s revenue and the utility of the
static agent (uti_1) are almost not affected by the type. So we draw the conclusion that the most
efficient way for the platform to earn more revenue is to attract more small companies to join the
mechanism, since in the figure 1 the revenue obviously increases as the number of agents increases.

6.2 Existence for Desirable Mechanisms

We assume all the agents’ type lies in [0, D], and the type space can be discretized into intervals
of length , which is also the minimal size of the data. Thus each agent’s type is a multiple of .
The data disparity is defined as the ratio of the largest possible data size to the smallest possible

13
 Welfare/Revenue/Utility
2.0
1.5 avg_wel avg_uti_1
avg_rev avg_uti_2
1.0
0.5
0.0
0 2 4 6 8 10
The type of opponent
Figure 2: Performance of MEP under different types

data size, denoted as, D/. We measure the condition for existence of desirable mechanisms by
the maximal data disparity when the market growth rate is given.
To describe the market growth, we use the following form of valuation function and model quality
function:
n
!α
X
Q(t) = t and vi (q) = Q(qj ) · Q(qi ), ∀i,
j=1

where the parameter α indicates the market growth rate. As we can see, when α < −1, the market
is not a growing market. When α ≥ 0, the market becomes non-competitive, therefore by Theorem
4, a desirable mechanism trivially exists. Thus we consider the competitive growing market case
where −1 ≤ α < 0.
We provide an algorithm to find the desirable mechanisms. Since the utility function is general, all
the points in the action space would influence the properties and existence of the mechanism, thus
the input space of the algorithm is exponential. The algorithm can be seen in Appendix A.1. We
enumerate the value of α and run the algorithm to figure out the boundary of D/ under different
α in a market with 2 agents. It turns out that when alpha is near −0.67, the boundary of disparity
grows very fast. When α is close to −0.66, the disparity boundary is over 10000 and is beyond our
computing capability, so we just enumerate α from −1 to −0.668 with step length 0.002.
Figure 3 demonstrates the maximal data disparity given the existence of a desirable mechanism
under different market growth rates. The y-axis is the data disparity D/ and the x-axis shows the
market growth rate α. For every fixed α, there does not exist any desirable mechanism when the
biggest size of the agent’s dataset is larger than D if the smallest size of dataset is . It can be seen
an obvious trend that when α becomes larger, the constraint on data size disparity would become
looser. In another word, if a desirable mechanism exists, a market with a larger α can tolerate a
larger data size disparity. A desirable mechanism is more likely to exist in a market that grows
faster. When the market is not growing, there would not be such a desirable mechanism at all. On
the other hand, if the market grows so fast such that there does not exist any competition between
the agents, the mechanism always exists.

14
 3k

2k

D/
1k

0
1.0 0.9 0.8 0.7
Market growth rate
Figure 3: The boundary of data disparity for existence of desirable mechanisms

Appendix

A Algorithm
A.1 Finding a Desirable Mechanism

We provide an algorithm, that given the agents’ valuations, computes whether the mechanism that
is simultaneously truthful, individually rational, efficient and weakly budget-balance exists, and
outputs the one that optimizes revenue, if any.
Since each agent can only under-report, according to the IR property, we must have:
ui (x(ti , t−i ), t) ≥ ui (x(∅, t−i ), t), ∀t, ∀i.
Equivalently, we get ∀t, ∀i,
ui (x(∅, t−i ), t) ≤ vi (x(ti , t−i ), t) − pi (ti , t−i ),
pi (ti , t−i ) ≤ vi (x(ti , t−i ), t) − ui (x(∅, t−i ), t),
pi (t) ≤ vi (x(ti , t−i ), t) − ui (x(∅, t−i ), t).
For simplicity, we define the upper bound of p(t0 ) as
p(t) , {vi (x(ti , t−i ), t) − ui (x(∅, t−i ), t).
The IC property requires that ∀ti ≥ t0i , ∀t−i , ∀i,
ui (x(ti , t−i ), t) ≥ ui (x(t0i , t−i ), t).
A little rearrangement gives:
pi (ti , t−i ) − pi (t0i , t−i )
≤vi (x(ti , t−i ), t) − vi (x(t0i , t−i ), t)
,Gapi (t0i , ti , t−i ).

15
Note that the inequality correlations between the payments form a system of difference constraints.
The form of update of the payments is almost identical to that of the shortest path problem. There-
fore, we make use of this observation to design the algorithm.
We assume that all the value functions are common knowledge, the efficient allocation is then
determined because the mechanism always chooses the one that maximizes the social welfare.
Thus it suffices to figure out whether there is a payment rule p(t0 ) which makes the mechanism IR,
IC and weakly budget-balance. Since the valid data size for each agent is bounded in practice, we
assume the mechanism only decides the payment functions on the data range [0, D], and discretize
the type space into intervals of length , which is also the minimal size of the data. Thus each
agent’s type is a multiple of . Note that since the utility function is general, all the points in the
action space would influence the properties and existence of the mechanism, thus it is necessary to
enumerate all the points in the space. The exponential value function space, i.e., the exponential
input space, determines that the complexity of our algorithm is exponential in D.
We give the following algorithmic characterization for the existence of a desirable mechanism.

Algorithm 1: Finding desirable mechanisms

input: Agents’ valuation functions v.
Use the function vi to calculate all the Gapi (t0i , ti , t−i ) and pi (ti , t−i ) for each i;
Initialize all pmax
i (ti , t−i ) to be pi (ti , t−i ) for each i;
for i = 1 to n do
for t−i = (∅, ∅, · · · , ∅) to (D, D, · · · , D) (increment =  on each dimension) do
Build an empty graph;
For each pi (ti , t−i ), construct a vertex Vti t−i and insert it into the graph;
Construct a base vertex V Bt−i which denotes the payment zero into the graph;
for ti = 0 to D (increment = ) do
Add an edge from V Bt−i to Vti t−i with weight p(ti , t−i );
for t0i = 0 to ti (increment = ) do
Add an edge with weight Gapi (t0i , ti , t−i ) from Vt0i t−i to Vti t−i ;
Use the Single-Source Shortest-Path algorithm to find the shortest path from V Bt−i to all
the vertices. These are the maximum solutions pmax
Pnothermax i (ti , t−i ) for each payment case;
if j=1 pj (t) < 0 then
return There is no desirable mechanism.
return pmax
i as the payment functions.

The following theorem proves the correctness of Algorithm 1.

Theorem 5. Taking agents’ valuation functions as input, Algorithm 1 outputs the answer of the
decision problem of whether there exists a mechanism that guarantees IR, IC, efficiency and weak
budget balance simultaneously, and specifies the payments that achieve maximal revenue if the
answer is yes.

Proof. Suppose that there is a larger payment for agent i such that pi (t0 ) > pmax
i (t0 ) where t0 is the
max 0
profile of reported types. In the process of our algorithm, the pi (t ) is the minimal path length

16
from V B−i to Vti t−i , denoted by (V B−i , Vti1 t−i , Vti2 t−i , · · · , Vtik =t0i t−i ). By the definition of edge
weight, we have the following inequalities:

pi (ti1 , t−i ) ≤ pi (ti1 , t−i ),

pi (ti2 , t−i ) − pi (ti1 , t−i ) ≤ Gapi (ti1 , ti2 , t−i ),
..
.
pi (tik , t−i ) − pi (ti(k−1) , t−i ) ≤ Gapi (ti1 , ti2 , t−i ).

Adding these inequalities together, we get

k−1
X
pi (t0 ) ≤ pi (ti1 , t−i ) + Gapi (tij , ti(j+1) , t−i ) = pmax
i (t0 ).
j=1

If pi (t0 ) < pmax

i (t0 ) holds, this would violate at least 1 of the k inequalities above. If the first
inequality is violated, the mechanism would not be IR, by the definition of pi (ti1 , t−i ). If any other
inequality is violated, the mechanism would not be IC, by the definition of Gapi (tij , ti(j+1) , t−i ).
On the other hand, if we select pmaxi (t0 ) to be payment of agent i, all the inequalities should be
satisfied, otherwise the shortest path would be updated to a smaller length.
Therefore the pmax
i (t0 ) must be the maximum payment for agent i. If the maximal payment sum up
to less than 0, there would obviously be no mechanism that is IR, IC and weakly budget-balance
under the efficient allocation function.

References
[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Tal-
war, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM
SIGSAC Conference on Computer and Communications Security, pages 308–318. ACM,
2016.
[2] Lawrence M Ausubel. An efficient ascending-bid auction for multiple objects. American
Economic Review, 94(5):1452–1475, 2004.
[3] Liad Blumrosen and Michal Feldman. Implementation with a bounded action space. In
Proceedings of the 7th ACM conference on Electronic commerce, pages 62–71. ACM, 2006.
[4] Liad Blumrosen and Noam Nisan. Auctions with severely bounded communication. In The
43rd Annual IEEE Symposium on Foundations of Computer Science, 2002. Proceedings.,
pages 406–415. IEEE, 2002.
[5] Shuchi Chawla, Hu Fu, and Anna R. Karlin. Approximate revenue maximization in interde-
pendent value settings. In Moshe Babaioff, Vincent Conitzer, and David A. Easley, editors,
ACM Conference on Economics and Computation, EC ’14, Stanford , CA, USA, June 8-12,
2014, pages 277–294. ACM, 2014.
[6] Michael Suk-Young Chwe. The discrete bid first auction. Economics Letters, 31(4):303–306,
1989.

17
 [7] Rachel Cummings, Stratis Ioannidis, and Katrina Ligett. Truthful linear regression. In Con-
ference on Learning Theory, pages 448–483, 2015.
[8] Esther David, Alex Rogers, Nicholas R Jennings, Jeremy Schiff, Sarit Kraus, and Michael H
Rothkopf. Optimal design of english auctions with discrete bid levels. ACM Transactions on
Internet Technology (TOIT), 7(2):12, 2007.
[9] Ofer Dekel, Felix Fischer, and Ariel D Procaccia. Incentive compatible regression learn-
ing. In Proceedings of the nineteenth annual ACM-SIAM symposium on Discrete algorithms,
pages 884–893. Society for Industrial and Applied Mathematics, 2008.
[10] Wenyi Fang, Pingzhong Tang, and Song Zuo. Digital good exchange. In Proceedings of the
2016 International Conference on Autonomous Agents & Multiagent Systems, pages 1277–
1278. International Foundation for Autonomous Agents and Multiagent Systems, 2016.
[11] Andrew V Goldberg and Jason D Hartline. Envy-free auctions for digital goods. In Proceed-
ings of the 4th ACM conference on Electronic commerce, pages 29–35. ACM, 2003.
[12] Andrew V Goldberg, Jason D Hartline, and Andrew Wright. Competitive auctions and digital
goods. In Proceedings of the twelfth annual ACM-SIAM symposium on Discrete algorithms,
pages 735–744. Society for Industrial and Applied Mathematics, 2001.
[13] Yaochen Hu, Di Niu, Jianming Yang, and Shengping Zhou. Fdml: A collaborative machine
learning framework for distributed features. In Proceedings of the 25th ACM SIGKDD In-
ternational Conference on Knowledge Discovery & Data Mining, pages 2232–2240. ACM,
2019.
[14] Jiawen Kang, Zehui Xiong, Dusit Niyato, Han Yu, Ying-Chang Liang, and Dong In Kim.
Incentive design for efficient federated learning in mobile networks: A contract theory ap-
proach. arXiv preprint arXiv:1905.07479, 2019.
[15] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Ar-
cas. Communication-efficient learning of deep networks from decentralized data. In Artificial
Intelligence and Statistics, pages 1273–1282, 2017.
[16] Reshef Meir, Ariel D Procaccia, and Jeffrey S Rosenschein. Strategyproof classification with
shared inputs. In Twenty-First International Joint Conference on Artificial Intelligence, 2009.
[17] Reshef Meir, Ariel D Procaccia, and Jeffrey S Rosenschein. Algorithms for strategyproof
classification. Artificial Intelligence, 186:123–156, 2012.
[18] Claudio Mezzetti. Mechanism design with interdependent valuations: Efficiency. Economet-
rica, 72(5):1617–1626, 2004.
[19] Paul R. Milgrom and Robert J. Weber. A theory of auctions and competitive bidding. Econo-
metrica, 50(5):1089–1122, 1982.
[20] Roger B Myerson and Mark A Satterthwaite. Efficient mechanisms for bilateral trading.
Journal of economic theory, 29(2):265–281, 1983.
[21] Javier Perote and Juan Perote-Pena. Strategy-proof estimators for simple regression. Mathe-
matical Social Sciences, 47(2):153–176, 2004.
[22] Ievgen Redko and Charlotte Laclau. On fair cost sharing games in machine learning. In
Thirty-Third AAAI Conference on Artificial Intelligence, 2019.

18
[23] Tim Roughgarden and Inbal Talgam-Cohen. Optimal and robust mechanism design with
interdependent values. ACM Trans. Economics and Comput., 4(3):18:1–18:34, 2016.
[24] Reza Shokri and Vitaly Shmatikov. Privacy-preserving deep learning. In Proceedings of the
22nd ACM SIGSAC conference on computer and communications security, pages 1310–1321.
ACM, 2015.
[25] Virginia Smith, Chao-Kai Chiang, Maziar Sanjabi, and Ameet S Talwalkar. Federated multi-
task learning. In Advances in Neural Information Processing Systems, pages 4424–4434,
2017.
[26] Ryo Yonetani, Vishnu Naresh Boddeti, Kris M Kitani, and Yoichi Sato. Privacy-preserving
visual learning using doubly permuted homomorphic encryption. In Proceedings of the IEEE
International Conference on Computer Vision, pages 2040–2050, 2017.

19