Attention-Enhanced CNN for Detecting AI-Generated Social Media Profile Images with Grad-CAM Visualization

Synthetic face images produced by modern generative AI systems have become nearly indistinguishable from genuine photographs, making them a preferred tool for building deceptive influencer personas on social media.Left unchecked, such profiles fuel misinformation campaigns, financial fraud, and a wider erosion of trust in online identity.Because manual review cannot scale to the volume of images uploaded every day, automated detection has become essential.This work presents a detection framework that pairs a ResNet-based convolutional neural network with a Convolutional Block Attention Module (CBAM), directing the model's focus toward the facial regions and channel features most diagnostic of synthetic origin.Training and evaluation relied on a balanced dataset of real and AI-generated face images drawn from public repositories, supplemented by rotation, flipping, and scaling to broaden generalization.Binary classification is performed through a sigmoid output, and Grad-CAM heatmaps are generated to expose which spatial cues drive each prediction.Experiments show the approach yields competitive detection performance while remaining interpretable-a combination that matters when deploying such tools in real-world content moderation pipelines.