Software-as-a-Service (SaaS) is a type of cloud computing in which a tenant rents access to a shared, typically web-based application hosted by a provider. Access control for SaaS should enable the tenant to control access to data that are located at the provider side, based on tenant-specific access control policies. Moreover, with the growing adoption of SaaS by large enterprises, access control for SaaS has to integrate with on-premise applications, inherently leading to a federated set-up. However, in the state of the art, the provider completely evaluates all policies, including the tenant policies. This (i) forces the tenant to disclose sensitive access control data and (ii) limits policy evaluation performance by having to fetch this policy-specific data. To address these challenges, we propose to decompose the tenant policies and evaluate the resulting parts near the data they require as much as possible while keeping sensitive tenant data local to the tenant environment. We call this concept policy federation. In this paper, we motivate the need for policy federation using an in-depth case study analysis in the domain of e-health and present a policy federation algorithm based on a widely-applicable attribute-based policy model. Furthermore, we show the impact of policy federation on policy evaluation time using the policies from the case study and a prototype implementation of supporting middleware. As shown, policy federation effectively succeeds in keeping the sensitive tenant data confidential and at the same time improves policy evaluation time in most cases.