SMILE. From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning - CVPR 2025