Network Function Virtualization (NFV) enables the implementation of Service Function Chains (SFCs), where a set of ordered Virtual Network Functions (VNFs) support diverse network services over resources. However, deploying SFCs across multiple domains faces challenges due to resource limitations and the confidentiality of domain-specific data. Achieving lowlatency SFC placement, followed by the deployed network service performance guarantee, is under research. Maintaining low endto-end (E2E) latency of network service delivery, while preserving the privacy of configuration details among Network Function Providers (NFPs) is also challenging the cross-domain SFC deployment. This paper focuses on reducing the E2E latency of cross-domain SFCs deployment while considering VNF scaling with data privacy. First, a Federated Learning with Graph Neural Network (FLGNN) approach is introduced among different NFPs to optimise VNF placement and SFC deployment while preserving data privacy. Then, a Viterbi-based SFC partitioning (VSFCP) method is proposed, which segments the SFC into sub chains based on FLGNN predictions, optimising E2E latency. After deploying the cross-domain SFC, a VNF auto-scaling method with transfer learning (TLVS) is presented to handle dynamic service demands. Results show that cross-domain SFC deployment latency decreases by up to 30.8%. VNF resource prediction error reduces by 39.3% to avoid service latency.