2026-08-132026-08-130026-06-14CANTÃO, Camily Furtado. Identificação de artefatos visuais em faces sintéticas por meio de transfer learning e IA explicável: uma abordagem investigativa com ResNet50 e Grad-CAM. Orientador: Iago Lins de Medeiros. 2026. 47 f. Trabalho de Curso (Bacharelado em Engenharia da Computação) – Faculdade de Engenharia da Computação, Campus Universitário de Tucuruí, Universidade Federal do Pará, Tucuruí, 2026. Disponível em: https://bdm.ufpa.br/handle/prefix/9923. Acesso em:.https://bdm.ufpa.br/handle/prefix/9923The advanced development of generative Artificial Intelligence technologies has led to the creation of widely accessible hyper-realistic synthetic content, solidifying deepfakes as a concrete threat to information integrity, privacy, and cybersecurity. This work investigates the use of Computer Vision and Explainable Artificial Intelligence (XAI) to identify visual features and structural artifacts associated with AI-generated sound images. To this end, a methodological pipeline was structured, composed of three macro-steps: curation of a balanced dataset of 10,000 real and synthetic material images, managed by the Roboflow platform; training of a binary classifier based on the ResNet50 architecture using Transfer Learning, executed on the Kaggle platform with NVIDIA Tesla T4 GPU support; and application of the Grad-CAM (Gradient-weighted Class Activation Mapping) algorithm to generate heatmaps that externalize the regions crucial for the model's decisions. As a preliminary exploratory study, an experiment was conducted with the YOLOv8 architecture, which revealed high computational cost and a lack of native spatial interpretability mechanisms, motivating the adoption of ResNet50 as the main methodological tool. The final classifier achieved an accuracy of 92.30%, an F1-Score of 92.29%, and an AUC-ROC of 0.9758, with no significant bias between classes. The heat maps generated by Grad-CAM revealed that the model directed its attention, in an anatomically coherent manner, to the central region of the face specifically the eyes, nose, and mouth in the synthetic images, regions consistent with the facial fusion artifacts described in the forensic literature. The results demonstrate that the c-ombination of Computer Vision and XAI offers not only high discriminative capacity but also transparency and auditability indispensable for digital forensics applications, overcoming the operational limitations of purely classificatory models.Acesso AbertoDeepfakesResNet50Transfer learningInteligência artificial explicávelGrad-CAMExplainable artificial intelligenceCNPQ::ENGENHARIASCNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAOTrabalho de CursoAttribution-NonCommercial-NoDerivs 3.0 Brazil