Behavioral Diagnostics of Deep Learning Models in Black-Box User Acceptance Testing Using a Prediction Model

The increasing adoption of deep learning models in modern software systems has significantly complicated User Acceptance Testing (UAT). Unlike conventional software, neural networks exhibit probabilistic behavior that depends on training data, model architecture, operational conditions, and robustness against adversarial perturbations. Most existing testing approaches rely on white-box or grey-box methodologies requiring access to gradients, neuron activations, or internal architectural information, which is usually unavailable during practical UAT. Consequently, there is a need for behavioral diagnostic methods capable of assessing deep learning model conditions using only externally observable characteristics. The purpose of this work is to develop a behavioral diagnostic approach and a prediction model for automated estimation of deep learning model conditions under black-box User Acceptance Testing constraints. Methodology. A generalized behavioral feature space was constructed using only externally observable metrics describing classification quality, prediction uncertainty, and adversarial robustness. Experimental data were generated from LeNet, VGG, ResNet, and DenseNet architectures trained on the CIFAR-10 dataset using FGSM, BIM, PGD, and DeepFool adversarial attacks. A custom Gated Multilayer Perceptron (Gated MLP) was implemented to perform multiclass classification of neural network operational conditions. The adaptive gating mechanism was used to estimate feature importance and improve prediction interpretability. The scientific novelty consists in developing a behavioral diagnostic method that performs automated assessment of deep learning model conditions under black-box constraints using only externally observable metrics. A generalized behavioral feature space combining quality indicators, uncertainty metrics, and adversarial perturbation characteristics was proposed. Feature importance analysis demonstrated that perturbation-related and uncertainty metrics provide the greatest contribution to diagnostic accuracy. Conclusions. Experimental evaluation demonstrated high prediction performance with an Accuracy of 0.9184 and an F1-macro score of 0.9175. Unlike existing white-box approaches, the proposed method does not require access to gradients, neuron activations, training data, or internal model architecture, making it suitable for practical User Acceptance Testing. The obtained results confirm the feasibility of automated behavioral diagnostics of deep learning systems using only externally observable characteristics.