Abstract
Implicitly harmful visual content represents a growing challenge in digital environments, particularly when offensive narratives are conveyed through stylized images, cartoons, or other indirect visual representations. Convolutional neural networks (CNNs) offer a promising approach for analyzing visual patterns associated with such content. This study presents an experimental framework based on transfer learning and cosine similarity to assess implicit harmful visual content under controlled conditions. Three pretrained CNN architectures, namely ResNet50, VGG19, and EfficientNetB0, were evaluated using accuracy, recall, F1-score, and confusion matrices. Rather than introducing a novel architecture, this work investigates the suitability of established CNN models combined with similarity-based feature representations as analytical instruments for visual content assessment. Cosine similarity was employed to examine relationships among deep feature embeddings and support the identification of subtle visual patterns. Experimental results demonstrated consistent performance across all evaluated architectures, with ResNet50 and VGG19 achieving the highest classification performance (Accuracy = 0.95; F1-score = 0.95), while EfficientNetB0 obtained competitive results with lower computational requirements. The proposed framework establishes a reproducible methodological basis for future studies involving visually sensitive content, semantic ambiguity, and the assessment of potentially offensive imagery generated by artificial intelligence systems, including text-to-image models.