Advancements in deep learning have enabled high-quality medical diagnostic services. Typically, high-tech corporations train deep learning models to provide these services to their clients. To access services, clients simply upload their medical data and receive predictions. However, due to the sensitivity of medical data, such services usually face privacy concerns from their clients. Additionally, service providers cannot directly deploy their models on clients’ devices, as the trained models are considered valuable intellectual property. To address this issue, secure multi-party computation (MPC) has been proposed as a solution, allowing service providers and clients to jointly perform deep learning computations without disclosing their respective inputs. For efficiency reasons, most MPC-based secure inference protocols assume that all parties will remain semi-honest throughout the execution of the protocols. However, this assumption may not always hold for clients. A recent study has shown that if a client is compromised by a malicious adversary, the adversary can steal the model’s parameters. However, such attack can be performed across limited neural network structures that only combines linear layers and ReLU layers. In this paper, we extend the attack to be applicable to most neural network structures commonly used in practice, e.g., VGG-Net and ResNet. Specifically, we design malicious steps for the adversary to circumvent max-pool and softmax layers. Our experimental results demonstrate that the extracted model parameters differ from the true values by at most 0.01.