Introduction
The red queen and the increasing non-transparency
Approach and methodology
Background and related work
Online trust and reputation systems
Visual analytics
Attacks and robustness of reputation systems
Attack | Description |
---|---|
Playbook | Playbooks are sequences of actions that allow the maximum outcome/profit for the player under specific auxiliary conditions. A simple example would be an on-off attack where a seller acts honestly to build high reputation by selling high quality products. After a particular time the seller changes his behavior and starts selling low quality products (under low production costs) and “milks” the high reputation. This sequence can be run through again and again. Overall, the seller stays unsuspicious [42]. This attack is particularly effective if the reputation metric “forgets” negative ratings. |
Value Imbalance Attack | Sellers making use of the value imbalance problem build up high reputation selling cheap products while cheating on the expensive ones. In contrast to playbooks, this attack is non periodical but the malicious seller has to keep a balance between good and bad behavior to keep his reputation at a certain level for the whole time. Zhang et al. [43] extended the term of value imbalance to transaction context imbalance where not only the value but also the product type or the time can be used for asymmetrical allocation of good service. |
Reputation Lag Attack | Usually, there is a time-lag between the advance payment and the delivery. Since referrals or ratings are normally made after the product is received, a malicious seller can exploit that time-frame by selling many low quality products before being rated badly for the first time. |
Proliferation Attack | In a proliferation attack, the seller offers the same product from several accounts or channels to increase the probability that a buyer chooses his product instead of buying from a different seller offering the same product. Although often named as an attack on reputation systems, the “malicious” sellers do not really cheat on their buyers nor do they manipulate the reputation system. However, they get an advantage of their competitors. (The proliferation attack can be considered as a subset of the Sybil attack) |
Re-entry Attack | Performing a re-entry attack (often referred to as whitewashing or newcomer attack), the malicious seller opens an account, cheats on the buyers and leaves the community to open a new account whenever his reputation is damaged. In re-entry attacks, the actor does never have to behave good. This type of attacks particularly exploits systems where a registration without any proof of identity is possible. |
Collusion | If multiple actors coordinate their behavior to gain an advantage over the rest, this is called collusion. The purpose of collusions can be various, e.g. unfairly increase/decrease the reputation of an actor, discriminate groups or run coordinated playbooks. |
Sybil Attack | In contrast to collusions where the accounts are created by multiple individuals, the Sybil attack is performed by one attacker who creates a number of accounts (pseudonyms). Due to the greater influence, the attacker can easily manipulate reputation values. |
Research gap: interactively visualizing seller reputation profiles to detect attacks
-
We want to increase the transparency of reputation systems through depicting all input data in integrated visual representations of the reputation profile (in TRIVIA)
-
We want to enhance the robustness of reputation systems through involving the user in the evaluation process (in TRIVIA)
Preliminary considerations for designing TRIVIA
Data classification
Information block | Data | Data type | Comment |
---|---|---|---|
Feedback | Multi-dimensional |
Typical feedback in electronic marketplaces involves both a rating and a textual review
| |
Rating | 1-dimensional | ||
Review | Text | ||
Time | 1-dimensional | ||
Transaction context | Multi-dimensional |
Product type, price and time as context attributes are exemplary chosen
| |
Product type | 1-dimensional | ||
Price | 1-dimensional | ||
Time | 1-dimensional | ||
Actors | Network |
Through direct ratings, a uni-directed referral graph is created
| |
Rater/advisor | Multi-dimensional | ||
Ratee | Multi-dimensional |
Visualization techniques
Feedback: rating
Feedback: review
Feedback: time
Transaction context: product type, price & time
Not yet rated transactions (reputation lag)
Interaction techniques
Data and view specification
View manipulation
Process and provenance
TRIVIA: a visual analytics tool to detect malicious sellers in electronic marketplaces
Conceptual design
Implementation
Case studies
The electronic market testbed
Case 1: The Playbook
Reputation system/model | Reputation value (after 100 days) |
---|---|
BRS | 0.91 |
Simple average | 0.6875 (84 % positive ratings) |
Case 2: value imbalance attack
Reputation system/model | Reputation value (after 100 days) |
---|---|
BRS | 0.86 |
Simple average | 0.90 (84 % positive ratings) |