<p>Violence remains a pervasive global challenge, causing severe physical, psychological, and emotional harm to individuals while destabilizing communities and straining societal resources. Its far-reaching consequences include a decline in public safety, economic losses, and long-term impacts on mental health, emphasizing the urgency of early detection and intervention. Automated violence detection systems, particularly those leveraging surveillance cameras, play a critical role in mitigating these threats, yet traditional methods often struggle under challenging conditions such as low-resolution inputs and dynamic environments. Moreover, the potential of graph-based approaches remains underexplored in this domain. Addressing these gaps, this paper introduces an innovative approach to detecting violent actions in video sequences by leveraging Graph Convolutional Networks (GCNs) and attention mechanisms. The proposed models, namely GCNonlyEdge, GCNVertexEdge, and the advanced SKE-A3TGCN, are designed to extract and learn features from graph-based representations of video data. Notably, the SKE-A3TGCN model excels by analyzing spatiotemporal characteristics of human skeleton movements through graph sequence data. This allows the model to focus on critical joints and their dynamic interactions, significantly enhancing its ability to identify violent behaviors in complex scenarios. Extensive experiments were conducted on benchmark datasets, including HockeyFight, RWF-2000, and Violence in Movies. The results highlight the superiority of the proposed models, with the SKE-A3TGCN model achieving the highest accuracy and demonstrating robust performance under challenging conditions. These findings underscore the effectiveness of the proposed solution in elevating the reliability of surveillance systems, contributing to enhanced public safety and security.</p>