Understanding and Improving Adversarial Attacks on Latent Diffusion Model