Violent behaviour poses a significant risk to societal stability and public safety. As part of proactive strategies to counteract this threat, many organizations and institutions have implemented surveillance systems to monitor and identify potential violent instances. Nevertheless, manual review and analysis of vast surveillance footage can be a daunting and error-prone task for human operators, necessitating the advent of automated systems for efficient and precise violence detection. This study introduces a novel approach for violence detection composed of a CNNRNN structure based on an attention mechanism for binary and multi-class classification of abnormal behaviours. We called our proposed model KianNet because Kian is the name of an intelligent innocent murdered in a violent incident, and we chose his name as a representative of all people who suffered from violent behaviours. In this technique, a CNN-RNN structure is applied to an input video to extract features from a sequence of frames and by adding a combination of Multi-Head Self-Attention (MHSA) and ConvLSTM layers, it can detect the violent event and determine the type of the observed anomaly. The key to KianNet’s performance is implementing the MHSA layer, which allows the model to focus on specific spatiotemporal regions of relevance, improving its capacity to differentiate between normal and violent events. Consequently, the MHSA layer boosts KianNet’s discriminatory power, enabling it to discern violent incidents from regular activities better. Through empirical evaluations, KianNet has proven its superior performance in violence detection tasks. Our findings reveal that KianNet outperforms its closest competitors’ accuracy by roughly 10 percent. This substantial performance margin substantiates the robustness and reliability of KianNet, cementing its potential as an effective tool in automated surveillance systems for violence detection.