Enhancing multi-corpus training in SSL-based anti-spoofing models: domain-invariant feature extraction