A Reputation System for Electronic Negotiations

Omid Tafreschi

, Dominique M

ahler

, Janina Fengel

2??

, Michael Rebstock

2??

and Claudia Eckert

Technische Univerti

at Darmstadt

Department of Computer Science

Darmstadt University of Applied Sciences

Faculty of Economics and Business Administration

Abstract. In this paper we present a reputation system for electronic negoti-

ations. The proposed system facilitates trust building among business partners

who interact in an ad-hoc manner with each other. The system enables market

participants to rate the business performance of their partners as well as the qual-

ity of offered goods. These ratings are the basis for evaluating the trustworthiness

of market participants and the quality of their goods. The ratings are aggregated

using the concept of Web of Trust. This approach leads to robustness of the pro-

posed system against malicious behavior aiming at manipulating the reputation

of market participants.

1 Introduction

Markets provide the basis for all types of business. According to [3], they have three

main functions, i.e., matchmaking between buyers and sellers, providing an institutional

infrastructure and the facilitation of transactions. A typical market transaction consists

of 5 phases [14], which are depicted in Figure 1:

Information

Phase

Intention

Phase

Agreement

Phase

Execution

Phase

Service

Phase

Fig. 1. Transaction Phases on a Market.

In the information phase market participants gather relevant information concerning

products, market partners, etc. During the intention phase market participants submit of-

fers concerning supply and demand. Then, the terms and conditions of the transaction

are speciﬁed and the contract is closed in the agreement phase. The agreed-upon con-

tract is operationally executed in the execution phase. During the service phase support,

maintenance and customer services are delivered.

The author is supported by the German Federal Ministry of Education and Research under

grant 01AK706C, project Premium.

These authors are supported by the German Federal Ministry of Education and Reserach under

grant 1716X04, project ORBI.

Tafreschil O., Mählerl D., Fengel J., Rebstock M. and Eckertl C. (2007).

A Reputation System for Electronic Negotiations.

In Proceedings of the 5th International Workshop on Security in Information Systems, pages 53-62

DOI: 10.5220/0002415300530062

 SciTePress

A negotiation can be deﬁned as a decentralized decision-making process by at least

two parties [5]. It is performed until an agreement is reached or the process is terminated

without reaching an agreement by one or all the partners involved. The objective of this

process is to establish a legally binding contract between the parties concerned, which

deﬁnes all agreed-upon terms and conditions next to regulations in case of failure of

their fulﬁllment [14]. Beam and Segev deﬁne electronic negotiations “as the process by

which two or more parties multilaterally bargain resources for mutual intended gain, us-

ing the tools and techniques of electronic commerce.” [4] Electronic negotiations enable

market participants to interact with each other in an ad-hoc manner over large geograph-

ical distances. However, these opportunities lead to undesirable side-effects. Over open

and anonymous networks, market participants have to cope with much higher mount of

uncertainty about the quality of products and the trustworthiness of other participants

[18]. Ackerlof showed that knowledge about the trustworthiness of a seller is vital for

the functioning of a market [1]. Therefore building trust in virtual environments is of

high importance. Information about the reputation can be used to build up trust among

strangers, e.g., new business partners. Learning from experiences from other users by

seeing their assessment of a potential partners’ business performance as well as the

quality of the delivered goods can reduce this uncertainty.

In this paper we propose a reputation system for electronic negotiations, which is

robust against unfair behavior. The remainder of this paper is organized as follows: In

Section 2, we present a business web supporting electronic negotiations, explain the

notion of reputation, and introduce reputation systems and their requirements. In Sec-

tion 3, we propose a reputation system and discuss its robustness against opportunistic

behavior. Section 4 presents related work. We conclude and discuss future work in Sec-

tion 5.

2 Background

2.1 A Business Web Forming an Electronic Market

The MultiNeg Project [8] has developed components for forming a business web which

shapes an electronic market. The market participants can offer their products, negoti-

ate with each other and transfer transaction results into their ERP systems. The busi-

ness web consists of four applications, i.e., marketplace, MultiNeg, Authentication and

Authorization Server (AAS) and ERP. Figure 2 shows the overall architecture of the

business web.

The marketplace allows sellers to publish catalogs containing their products or ser-

vices, which can be subject to negotiations.

MultiNeg is an electronic negotiation support system for bilateral, multi-attributive

negotiations. It facilitates the negotiation of multiple items with arbitrary variable at-

tributes. The key objectives for the development have been the design of an architecture

suitable for different industries, company sizes and products and a communication inter-

face design that allows the integration of inter-organizational with intra-organizational

applications. The functionalities are conceptualized for usage in a decentralized deploy-

ment and are based on open standards, e.g., HTTP, XML, SOAP, WSDL. Thus they

allow the seamless electronic integration of internal and external business processes.

MARKETPLACE

MULTINEG

1.) publish catalogue

3.) initialize negotiation

BUYER

AAS

2.) search products

ERP

6.) transmit results

SELLER

AAS

5.) finish negotiation

4.) negotiate products

ERP

6.) transmit results

5.) finish negotiation

4.) negotiate products

Fig. 2. Architecture of Business Web.

Each company has an Authentication and Authorization Server (AAS). This allows

market participants to setup security policies for authentication and authorization. Au-

thentication policies deﬁne how business partners of a company can authenticate them-

selves in order to interact with the owner of the AAS. Authorization policies support

the implementation of access control to the resources of a company, e.g., catalog data.

As mentioned in the introduction, trust building is crucial for electronic negotia-

tions. Information about reputation can be used for that purpose. In the following we

describe the notion of reputation.

2.2 The Notion of Reputation

According to [17], reputation is a concept that arises in repeated game settings when

there is uncertainty about some property of one or more players in the mind of other

players. If “uninformed” players have access to the history of past stage game outcomes,

reputation effects then often allow informed players to improve their long-term payoffs

by gradually convincing uninformed players that they belong to the type that best suits

their interests.

Reputation information is useful in two respects. First, it supports a market partic-

ipant with regard to decision support. This is necessary when a buyer has to choose

between several sellers offering the same product with comparable conditions. Second,

it makes opportunistic behavior unproﬁtable, since this would lead to a poor rating,

which in turn may deter potential business partners from considering the owner of the

poor rating.

2.3 Reputation Systems and their Requirements

A reputation system collects, distributes and aggregates feedback about participants’

past behavior [15]. Feedback is usually collected at the end of each transaction. The au-

thority responsible for reputation management asks the participants to submit a rating.

Aggregation of feedback means that the available information about an entity’s trans-

actions is evaluated and condensed into a few values that allow users to easily make

a decision. Feedback is distributed ﬁnally to the participants, i.e., it has to be made

available to everyone who wants to use it for making decisions on whom to trust.

Figure 3 shows the components and actors of a reputation system [16]. The target

of a rating is called ratee. The collector gathers ratings from agents called raters. This

information is processed and aggregated by the processor. The emitter makes the results

available to other requesting agents.

Collector Processor Emitter

Rater

Evaluator

Ratee

4.) update

)

2.) interaction

Fig. 3. Components and Actors of Reputation Systems [16].

The main objective of a reputation system is to prevent opportunistic behavior in

virtual environments by making it worthless in the long term. However, since the same

communication facilities are used for implementing reputation systems, these systems

face themselves some problems caused by opportunistic behavior of market participants

using reputation systems.

In [9], Dellarocas presents two scenarios in which market participants try to ma-

nipulate their own reputation or the reputation of others, i.e., ballot stufﬁng and bad-

mouthing. Ballot stufﬁng is the result of unfairly high ratings, which are given by a

colluded group of buyers to a single seller in order to improve his reputation. Bad-

mouthing describes a situation in which colluded buyers give unfairly negative ratings

to a seller in order to drive him out of the market.

In [10], Friedman and Resnick describe the dilemma of cheap pseudonyms. This

occurs when a user of an online service can change his virtual identity by obtaining

a new pseudonym at little cost. As a consequence, market participants can easily get

rid of negative reputations. Such market participants are referred to as whitewashers:

users who leave a system and rejoin with new identities to avoid reputational penalties.

Cheap pseudonyms allow another adversarial strategy, i.e., sybil attacks [7]. Sybils are

different identities of one user. This user improves his reputation by using his sybils.

In addition to the issues mentioned above, reputation systems supporting mutual

ratings should prevent that one party can learn the rating he is supposed to receive

before he has issued the rating for his business partner. This should ensure that a rating

reﬂects its issuer’s experience in the context of the corresponding transaction and that a

rating is independent from the rating of the opposite side.

3 A Reputation System for Electronic Negotiations

In this section we present a reputation system for the business web mentioned above. It

enables market participants, i.e., buyers and sellers, to rate each other after a concluded

transaction. They can assess the quality of the traded goods as well as the partner’s

performance. This mechanism enables participants to make better decisions considering

several aspects. For example a seller is offering products of excellent quality, but tends

to deliver late. This information is valuable for potential buyers who have to plan their

production in advance. If they decide to buy some goods from the seller in question,

they have to schedule reserve time for maybe late delivery.

Our reputation system encompasses two phases, i.e., rating phase and rating ag-

gregation phase. In the rating phase two business partners evaluate each other after a

concluded transaction by giving a rating to the opposite side. This phase affects the last

three transaction phases on a market during which two business partners interact with

each other. During these three phases including the negotiation they form an opinion

about each other, which is later reﬂected by the corresponding ratings. These ratings

are used in the rating aggregation phase to get information about the reputation of a

market participant. The rating aggregation phase takes place during the ﬁrst three trans-

action phases. In the information phase and the intention phase a buyer wants to learn

more about his potential negotiation partners, i.e., sellers. For this purpose he has to

evaluate the rating of sellers. During the actual negotiation both buyer and seller can

use ratings of their opposite to map out their negotiation strategy. Figure 4 depicts the

rating phase and the rating aggregation phase during a market transaction.

Marketplace MultiNeg ERP

Information

Phase

Intention

Phase

Agreement

Phase

Execution

Phase

Service

Phase

Rating Phase

Rating Aggregation Phase

Fig. 4. Rating Phase and Rating Aggregation Phase During a Market Transaction.

3.1 Rating Phase

Figure 5 shows the process of a negotiation and the subsequent rating phase. We de-

scribe the process in the following:

Steps 1-2 After concluding a negotiation the involved parties evaluate each other

by issuing corresponding ratings. A rating ranges from 0, i.e., poor rating, to 100, i.e.,

excellent rating. We support two different types of rating, i.e., buyer rating and seller

rating. A buyer rating is given by a seller to a buyer, whereas a seller rating is given

by a buyer to a seller. Each kind of rating has a set of speciﬁc categories, which can

be evaluated separately. For example, a seller can rate the payment practice of a buyer

and a buyer can rate the quality of products delivered by a seller. A weighting can be

assigned to each category to reﬂect the importance of this category to the evaluator. The

sum of all weightings of a single rating is 100. The default value of a weighting is 100

divided by the number of the categories. Additionally, a category can be annotated with

a comment. Since the data of a negotiation are stored on MultiNeg, negotiation partners

use MultiNeg to create ratings for each other.

MARKETPLACE

MULTINEG

2.) issue rating

6.) verify signature

1.) conclude negotiation

7.) check the

existence of the

transaction

SELLER

AAS

2.) issue rating

1.) conclude negotiation

8.) publish ratings

4.) sign rating

AAS

4.) sign rating

BUYER

ERP ERP

)

Fig. 5. Rating Phase.

Steps 3-5 After the receipt of the rating data MultiNeg forwards it to the AAS of

the corresponding party. The AAS displays the rating to its issuer and allows him to

digitally sign it. The signed rating is transferred to the marketplace to be published.

Steps 6-8 The signature of a signed rating is veriﬁed at the marketplace. In case

of a positive result the marketplace checks whether the rating belongs to an existing

negotiation. This step ensures that only authentic ratings are published. If MultiNeg’s

answer is positive, the marketplace publishes the rating if both ratings of a negotiation

have been issued. Publishing two ratings simultaneously prevents that one negotiation

partner changes his mind (rating) after learning the rating given to him by his opposite.

However, this could protect one party from receiving a poor rating. In this case the party

does not give a rating for his opposite. To overcome this shortcoming the marketplace

publishes a single rating without the second corresponding rating after a deﬁned period

of time has passed. This rule is known to all market participants.

3.2 Rating Aggregation Phase

Existing ratings of market participants are stored and displayed at the marketplace

where buyers search the catalogs of sellers and choose partners for a negotiation. The

ratings of market participants can be used to support that decision. Our reputation sys-

tem provide a set of different views to the ratings. An evaluator can look at details of

one rating, e.g., its categories and the corresponding comments. In order to simplify

the assessment of a certain market participant based on his ratings, our system allows

calculating the arithmetic average of all ratings, i.e., ratings, of one market participant

with the help of the following formulas. We assume that the market participant has

n ratings and each rating consists of m categories.

rating =

j=1

category

· weighting

100

(1)

ratings

l=1

rating

(2)

Formula 1 calculates the average of one rating. We apply this to all ratings of mp

and calculate the average of them with Formula 2. Although this approach is easy to

implement and to understand, it is not robust against ballot stufﬁng and bad-mouthing,

since it considers only the number of ratings independently from their issuers. In the

following we show how to overcome this shortcoming.

If we assume that trust is transitive, we can apply the concept of Web of Trust for the

calculation of trust values based on rating lists. Pretty Good Privacy (PGP) was the ﬁrst

popular system which used the concept of Web of Trust to establish the authenticity of

the binding between a public key and a user in a decentralized and open environment,

e.g., the Internet [19]. We use the concept of Web of Trust to calculate trust values in our

reputation system. For this purpose, the evaluator ﬁrst has to build up a mesh between

himself and the peer for which he wants to calculate the trust value. The mesh should

consist of one trust path at least. The start vertex of a trust path is the evaluator and the

end vertex is the market participant to be evaluated. In addition, a trust path contains

other vertices, i.e., market participants which have been rated by the evaluator or by

other vertices in the mesh except the end vertex. In order to keep the calculation base

meaningful, we limit the length of a trust path, i.e., the number of its vertices [11]. The

average of the rating which was given to a vertex is calculated with Formula 1. If there

are several ratings, i.e., weightings to be assigned to one edge, we aggregate them with

Formula 2. The result is assigned as weighting to the edge directing to the vertex in

question. It is not always possible to ﬁnd a trust path, e.g., when no in between vertices

exist. Our reputation system indicate such cases to the evaluator.

We explain the algorithm for the aggregation of ratings based on Web of Trust with

the help of an example. One buyer (Company A) enters the marketplace and wants

to purchase products which are offered by the seller (Company Z). These two market

participants have never negotiated with each other. But there are two other companies

which have negotiated with both, seller and buyer. The corresponding ratings are shown

in Figure 6. We use the existing ratings to build up a mesh between buyer and seller. If

Rater Ratee Rating

Company A (buyer) Company B 95%

Company A (buyer) Company C 70%

Company A (buyer) Company C 90%

Company B Company Z (seller) 99%

Company C Company Z (seller) 95%

start node

target node

Fig. 6. Example for Web of Trust.

there are several ratings between two parties we aggregate them into one value, which

is the arithmetic average. Figure 6 shows the corresponding mesh of our example. The

mesh is the input for the following algorithm proposed by Caronni [6].

Algorithm 1: calculateTrust.

Input: mesh W, start node A, target node Z

Output: TrustLevel

trust in A = 1;1

TrustLevel = 0;2

for all direct neighbors of Z do3

direct neighbor is X;4

if X is the start node then5

T rustLevel = 1 − (1 − (trust in X) · (edge

X,Z

) · (1 − T rustLevel);6

else7

create a copy C of W;8

drop target node Z in C;9

calculateTrust(C,A,X);10

T rustLevel = 1 − (1 − return · edge

X,Z

) · (1 − T rustLevel);11

return T rustLevel;12

The algorithm is recursive and has three runs for our example mesh. The direct neigh-

bors from Z are node B and node C. The intermediate steps are depicted in Figure 7.

The algorithm starts with node B, which is not the start node. As a consequence, in the

left mesh the company Z (seller) has been removed and company B is the new target

node (see Figure 7(a)). We ﬁnd out the direct neighbors of node B, which is node A

(start node). Line 6 of the algorithm contains the formula for calculating the trust level

of node A with regard to node B (T rustLevel

A,B

= 1− (1− 1· 0, 95)· (1− 0) = 0, 95).

This value is returned from the recursive call. It is the input in the formula in line 11

of the algorithm, where we calculate the TrustLevel of node A with regard to node Z

(T rustLevel

A,B,Z

= 1 − (1 − 0, 95 · 0, 99) · (1 − 0) = 0, 9405).

start node

target node

(a)

start node

target node

(b)

Fig. 7. Iteration Steps for Calculating a Trust Level.

We repeat the sketched process again for node C as next neighbor of node Z (see

Figure 7(b)). The resulting value is T rustLevel

A,C

= 1 − (1 − 1 · 0, 8) · (1 − 0) = 0, 8.

In the ﬁnal step T rustLevel

A,B,Z

and T rustLevel

A,C

are the inputs for the formula

in line 11 where the return value of the algorithm is calculated T rustLevel

A,Z

1 − (1 − 0, 8 · 0, 95) · (1 − 0, 9405) = 0, 98572. This value is presented to Company A

in addition the arithmetic average of Company B’s ratings. This additional information

provides an alternative for decision-making considering the perspective of the evaluator.

The aggregation of ratings based on the concept of Web of Trust does not totally

prevent ballot stufﬁng and bad-mouthing, but these kinds of malicious behavior lose

some of their impact. Recall that ballot stufﬁng requires that a (small) set of raters give

a large number of unfair good ratings to a ratee in order to improve his reputation. Since

the approach based of Web of Trust considers the average of ratings given to a market

participant, it does defuse the impact of several unfair ratings. The same reasoning

applies to bad-mouthing.

4 Related Work

Au et al. present a framework for enabling crossorganisational trust establishment [2].

They introduce trust tokens which are issued by trust servers. These tokens are required

when a user wants to access a protected resource. Au et al. also use the concept of Web

of Trust to model the trust relationship between the trust servers. However, in contrast

to their work, we build up a Web of Trust based on ratings and can calculate the trust

level in an ad-hoc manner.

Kamvar et al. propose the EigenTrust algorithm for reputation management in P2P

networks [12]. This work is close to our system, since both are based on the concept of

transitive trust. But they have different design goals. EigenTrust minimizes the impact

of malicious peers on the performance of a P2P system.

In [13], Maurer proposes an approach to model a user’s view of a public key infras-

tructure (PKI). From this view, a user draws conclusions about the authenticity of other

entities’ public keys and possibly about the trustworthiness of other entities. A user’s

view consists of the user’s statements about the authenticity of public keys and the trust-

worthiness of their owners, as well as a collection of certiﬁcates and recommendations

obtained or retrieved from the PKI. This approach is similar to our work. However, we

use other information, i.e., transaction ratings, to calculate the trustworthiness of market

participants.

5 Conclusion and Future Work

In this paper, we proposed a reputation system for electronic negotiations to enable

market participants to evaluate each other and their offered goods. For this purpose, we

developed a system architecture which allows business partners to evaluate each other

with the help of ratings after a concluded transaction. These ratings are aggregated to

support a market participant to ﬁnd a trustworthy business partner. This aggregation is

based on the concept of Web of Trust to make the reputation systems robust against

malicious behavior aiming at manipulating the reputation of some market participants.

In addition to the aggregated values, our approach offers differentiated views for

assessing a business partners’ capabilities. Instead of an overall information, a potential

partner can get insight about the business behavior and the quality of the offered goods

separately, since those do not necessarily depend on each other.

We have implemented the proposed reputation system and plan to do an empirical

study of the usability of our system and to analyze its efﬁciency.

References

1. G. A. Akerlof. The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism.

Quartely Journal of Economics, 84(3):488–500, 1970.

2. R. Au, M. Looi, and P. Ashley. Automated Cross-Organisational Trust Establishment on

Extranets. In ITVE ’01: Proceedings of the Workshop on Information Technology for Virtual

Enterprises, pages 3–11, Washington, DC, USA, 2001. IEEE Computer Society.

3. Y. Bakos. The Emerging Role of Electronic Marketplaces on the Internet. Communincations

of the ACM, 41(8):35–42, 1998.

4. C. Beam and A. Segev. Automated Negotiations: A Survey of the State of the Art.

Wirtschaftsinformatik, 39(3):263–268, 1997.

5. M. Bichler. A Roadmap to Auction-based Negotiation Protocols for Electronic Commerce.

In HICSS ’00: Proceedings of the 33rd Hawaii International Conference on System Sciences,

volume 6, Washington, DC, USA, 2000. IEEE Computer Society.

6. G. Caronni. Walking the Web of Trust. In WETICE ’00: Proceedings of the 9th IEEE

International Workshops on Enabling Technologies, pages 153–158, Washington, DC, USA,

2000. IEEE Computer Society.

7. A. Cheng and E. Friedman. Sybilproof Reputation Mechanisms. In P2PECON ’05: Pro-

ceeding of the 2005 ACM SIGCOMM Workshop on Economics of Peer-to-Peer Systems,

pages 128–132, New York, NY, U SA, 2005. ACM Press.

8. The MultiNeg Consortium. http:www.fbw.h-da.de/multineg, 2007.

9. C. Dellarocas. Immunizing Online Reputation Reporting Systems Against Unfair Ratings

and Discriminatory Behavior. In EC ’00: Proceedings of the 2nd ACM Conference on Elec-

tronic commerce, pages 150–157, New York, NY, USA, 2000. ACM Press.

10. E. Friedman and P. Resnickl. The Social Cost of Cheap Pseudonyms. Journal of Economics

& Management Strategy, 10(2):173–199, 2001.

11. A. Jøsang, R. Hayward, and S. Pope. Trust Network Analysis with Subjective Logic. In

ACSC ’06: Proceedings of the 29th Australasian Computer Science Conference, pages 85–

94, Darlinghurst, Australia, Australia, 2006. Australian Computer Society, Inc.

12. S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina. The Eigentrust Algorithm for Repu-

tation Management in P2P Networks. In WWW ’03: Proceedings of the 12th international

conference on World Wide Web, pages 640–651, New York, NY, USA, 2003. ACM Press.

13. U. M. Maurer. Modelling a Public-Key Infrastructure. In E. Bertino, H. Kurth, G. Martella,

and E. Montolivo, editors, Computer Security - ESORICS 96, 4th European Symposium on

Research in Computer Security, volume 1146 of Lecture Notes in Computer Science, pages

325–350. Springer, 1996.

14. M. Rebstock and M. Lipp. Webservices zur Integration interaktiver elektronischer Verhand-

lungen in elektronische Marktpl

atze. Wirtschaftsinformatik, 45(3):293–306, 2003.

15. P. Resnick, K. Kuwabara, R. Zeckhauser, and E. Friedman. Reputation systems. Communi-

cations of the ACM, 43(12):45–48, 2000.

16. A. Schlosser, M. Voss, and L. Br

uckner. On the Simulation of Global Reputation Systems.

Journal of Artiﬁcial Societies and Social Simulation, 9(1), 2005.

17. R. Wilson. Reputations in Games an Markets. Game Theoretic Models of Bargaining, 1985.

18. W. Zhang, L. Liu, and Y. Zhu. A Computational Trust Model for C2C Auctions. In Ser-

vices Systems and Services Management, Proceedings of ICSSSM ’05. 2005 International

Conference, 2005.

19. P. R. Zimmermann. The ofﬁcial PGP User’s Guide. MIT Press, Cambridge, MA, USA,

1995.