A Learning-based Approach to Information Release Control

Controlled release of information from an organization is becoming important from various considerations: privacy, competitive information protection, strategic data control, and more. In most organizations, data protection is afforded only by using access control. However, it can be argued that access control suffers from at least two problems. First, effective access control assumes a perfect categorization of information ("who can access what"), which is increasingly difficult in a complex information system. Second, access control is not effective against insider attacks, where users with legitimate access rights send out sensitive information, either with malicious intent or by accident. Information release control is viewed as complementary to access control, and aims at restricting the outgoing information flow at the boundary of information systems. This paper presents an architectural view of a release control system. The system emphasizes the role of automated learning for release control constraints. This has resulted from the realization that the most difficult task of effective release control is how the release control constraints are specified. In a learning-based system, data mining and machine learning techniques are employed to generate release control constraints from samples provided by the security officer. The system applies continuous learning to adjust the release control constraints to reduce both mistakenly released and mistakenly restricted documents. This paper also provides a specific example on how to learn keyword-based release control constraints.