A Review on Deep Transfer Learning Techniques for Detecting Abusive and Harmful Visual Content

Abstract

The rapid expansion of digital communication platforms has led to an unprecedented increase in the sharing of visual content, making the automatic identification of abusive and harmful images a critical research challenge. Harmful visual materials—including violent, hateful, explicit, and disturbing images—pose significant risks to online communities, emphasizing the need for intelligent and scalable content moderation systems. Deep transfer learning has emerged as a highly effective approach for this task by utilizing knowledge acquired from large-scale pre-trained models and adapting it to domain-specific datasets with limited labeled samples. This review presents a comprehensive analysis of recent deep transfer learning techniques employed for abusive visual content detection. It examines widely adopted architectures, including Convolutional Neural Networks (CNNs), EfficientNet, MobileNet, ResNet, Vision Transformers (ViTs), hybrid CNN–Transformer models, and multimodal vision–language frameworks. The survey also reviews benchmark datasets, transfer learning strategies, evaluation metrics, and comparative performance reported in recent studies. Furthermore, it discusses key challenges such as dataset scarcity, class imbalance, cultural bias, adversarial robustness, interpretability, computational complexity, and privacy concerns that continue to affect the reliability of automated moderation systems. Emerging research directions—including explainable artificial intelligence, few-shot and zero-shot learning, multimodal fusion, foundation models, and privacy-preserving learning—are also highlighted as promising solutions for developing more accurate, robust, and trustworthy content moderation frameworks. This review provides researchers and practitioners with a structured overview of current developments, identifies existing research gaps, and outlines future opportunities for advancing deep transfer learning techniques in the detection of abusive and harmful visual content.