Public Dataset and Explainable AI for Cognitive Cybersecurity: Grad-CAM Insights into Negative Visual Stimuli Detection
Konstantin Olegovich Gnidko, Dmitry Lisov · 2025
The rapid proliferation of digital content has raised concerns regarding the detection and mitigation of harmful visual stimuli, particularly in cognitive cybersecurity. This study introduces a novel framework for detecting destructive visual content using a convolutional neural network (CNN) based on the ResNet50 architecture, fine-tuned for binary classification problem. A comprehensive dataset was curated, integrating subjective annotations from participants and established image repositories, and refined through anomaly detection via the isolation forest algorithm. The CNN was trained and validated using k-fold cross-validation, achieving high precision and recall rates despite the inherent variability in emotional responses. To enhance interpretability, the Grad-CAM visualization technique was applied, allowing critical analysis of the model’s decision-making process. The findings underscore the potential of integrating advanced deep learning techniques with innovative anomaly detection to address challenges in recognizing harmful multimedia content. This work lays the groundwork for future research into adaptive and personalized models for diverse cognitive environments, contributing to the broader field of automated content moderation and cybersecurity.