Title: ARCHITECTURE FOR PRIVACY-PRESERVING INFORMATION FILTERING

This section describes an architecture for privacy-preserving information filtering based on the employment of a TCG-compliant platform on the server side. The solution addresses the use case described in section , but in addition to preserving the privacy of the user data, it takes into account the privacy of the other actors, namely the information service provider and the recommender system provider, resulting in an architecture which preserves privacy in a multilateral way. The architecture is based on Multi-Agent System (MAS) technology because fundamental features of agents such as autonomy, adaptability and the ability to communicate are essential requirements of the chosen approach.

Definitions and Requirements

There are three main abstract entities participating in an information filtering process within a distributed system: A user entity, a provider entity and a filter entity. Whereas in some applications the provider and filter entities explicitly trust each other, because they are deployed by the same party, the described solution is applicable more generically because it does not require this kind of explicit trust between the main abstract entities.  

The user is the entity which intends to obtain personalized recommendations, based on private data collected in a user profile. The information these recommendations are based on is collected in the provider profile linked to the information provider entity. The filter entity provides filtering techniques, i.e. the algorithms used to generate the recommendations. The following sections focus on aspects related to the information filtering process itself, and omit all aspects related to information collection and processing, i.e. the stages in which profiles are generated and maintained, mainly because these stages are less critical with regard to privacy, as they involve fewer different entities. 

The architecture aims at meeting the following requirements with regard to privacy: 

1. User Privacy: No linkable information about user profiles should be acquired permanently by any other entity or external party, including other user entities. Single user profile items, however, may be acquired permanently if they are unlinkable, i.e. if they cannot be attributed to a specific user or linked to other user profile items. Temporary acquisition of private information is permitted as well. Sets of recommendations may be acquired permanently by the provider, but they should not be linkable to a specific user. These concessions simplify the resulting protocol and allow the provider to obtain recommendations and single unlinkable user profile items, and thus to determine frequently requested information and optimize the offered information accordingly.

2. Provider Privacy: No information about provider profiles, with the exception of the recommendations, should be acquired permanently by other entities or external parties. Again, temporary acquisition of private information is permitted. Additionally, the propagation of provider information is entirely under the control of the provider. Thus, the provider should be enabled to prevent misuse such as the automatic large-scale extraction of information.

3. Filter Privacy: Details of the algorithms applied by the filtering techniques should not be acquired permanently by any other entity or external party. General information about the algorithm may be provided by the filter entity in order to help other entities to reach a decision on whether to apply the respective filtering technique.

Additionally, general requirements regarding the quality of the recommendations as well as security aspects, performance and broadness of the resulting system have also to be addressed. While minor trade-offs may be acceptable, the resulting system should reach a level similar to existing (non-privacy-preserving) recommender systems with regard to these requirements.