@article{PANG2026, 
author = {Meng PANG and Yintao ZHOU and Rong ZHANG and Jingjing ZHANG and Nanrun ZHOU},
title = {Comprehensive teaching experiment on face prototype reconstruction based on adversarial attack defense},
year = {2026},
journal = {Experimental Technology and Management},
volume = {43},
number = {4},
pages = {179-186},
keywords = {comprehensive teaching experiment, experimental teaching case, teaching reform, face prototype reconstruction, generative adversarial networks, adversarial attack defense},
url = {https://www.sciopen.com/article/10.16791/j.cnki.sjg.2026.04.022},
doi = {10.16791/j.cnki.sjg.2026.04.022},
abstract = {ObjectiveCurrent experimental teaching in the computer vision field faces challenges such as outdated content and limited integration of cutting-edge research into talent cultivation. These limitations are particularly problematic in courses designed for senior students who seek practical research experience. To address this issue, this paper designs a teaching experiment on face prototype reconstruction based on adversarial attack defense. The proposed experiment has significant pedagogical value because it addresses the real-world problem in which face recognition fails for images affected by variations such as pose, expression, and illumination, while also incorporating the critical issue of AI security. The experiment aims to guide students systematically through the entire process of model design, training, and testing to reconstruct standardized, identity-preserved face prototypes from contaminated images, while simultaneously learning the design and implementation of adversarial defense strategies.MethodsThe method used in the proposed experiment consists of two core components: a prototype reconstruction generative adversarial network and an adversarial defense network. Specifically, (1) the prototype reconstruction generative adversarial network comprises an encoder–decoder structured generator and a multi-task discriminator. It is designed to reconstruct variation-free face prototypes by jointly learning prototypes and representations within a unified framework. The network is optimized through an adversarial game between the generator and the discriminator. As the discriminator improves its ability to perform identity prediction, facial variation detection, and real–fake prototype classification, the generator simultaneously learns to produce realistic prototype images that preserve identity information while removing facial variations, thereby deceiving the discriminator. (2) The adversarial defense network aims to mitigate the adverse effects of adversarial perturbations by preprocessing the input images of the prototype reconstruction generative adversarial network through a combination of image denoising and matrix estimation. The network includes four key modules: a feature enhancement module for extracting latent features, an attention module that refines noise-related features using an attention mechanism, a residual learning module that subtracts the predicted noise from the original input, and a matrix estimation module employing Universal Singular Value Thresholding (USVT) to restore the global image structure. For the experimental procedure, students are guided through three stages. First, data preparation is conducted by dividing facial images under visible light in the LAMP-HQ dataset into training and testing sets. Second, model training is performed for both the prototype reconstruction generative adversarial network and the adversarial defense network, during which students are encouraged to adjust hyperparameters to better understand the training process. Third, model testing is conducted by generating adversarial samples to evaluate the effectiveness and robustness of prototype reconstruction under adversarial attacks.ResultsThrough both qualitative and quantitative evaluation, students can verify the effectiveness of the proposed method based on the experimental results. In the qualitative evaluation, visualization of reconstruction results demonstrates the effectiveness of the prototype reconstruction generative adversarial network in generating identity-preserved prototypes from contaminated inputs, as well as the effectiveness of the adversarial defense network in restoring prototype image quality that may otherwise be significantly degraded by adversarial attacks. In the quantitative evaluation, objective metrics including the Rank-10 recognition rate and image sharpness measured by the Laplacian variance method are used to further assess performance. The results confirm the negative impact of adversarial attacks on prototype reconstruction, with the Rank-10 recognition rate decreasing by 22.51%, 31.86%, and 38.25% at perturbation magnitudes of 0.15, 0.20, and 0.25, respectively, while image sharpness exhibits a similar downward trend. After integrating the adversarial defense network, the Rank-10 recognition rate recovers by 7.61%, 14.64%, and 15.86% at the corresponding perturbation magnitudes, and image sharpness demonstrates a similar upward trend.ConclusionsThis paper develops a comprehensive teaching experiment that translates advanced research findings in computer vision and AI security into a feasible pedagogical case. The experiment bridges the gap between theory and complex engineering challenges by guiding students through model design, implementation, and rigorous experimental analysis for face prototype reconstruction based on adversarial attack defense. Teaching practice demonstrates that the experiment helps students transition from passive verifiers to active problem-solvers by improving their engineering skills, empirical research abilities, and innovative thinking, while simultaneously fostering awareness of AI security. The proposed experiment therefore provides a valuable and replicable paradigm for reforming computer vision curricula.}
}