Nowadays, the number of cyberattacks that affect machine learning and deep learning models is increasing. These models may be vulnerable to different attacks, including adversarial attacks that can compromise information at various stages of data processing. For example, the model inversion attack is an exploratory attack that focuses on compromising the information privacy, recovering the training data, or inferring sensitive information from these data. In this type of attack, multiple loss functions have been employed in the literature, including Cross-Entropy loss, Max-Margin loss, and Poincaré loss. The objective of this paper is to study the stability and effectiveness of these loss functions in the training process of the selected Model Inversion attack method. In addition, the number of iterations required to achieve high attack performance is analyzed, taking into account the loss function. To this end, target models with varying degrees of complexity and datasets of different sizes are utilized.