DiffMI: Breaking Face Recognition Privacy via Diffusion-Driven Training-Free Model Inversion