As internet technology continues to advance and become more widely used, the number of cyberattacks is on the rise. One of the most damaging forms of attack is the botnet. The proliferation of attack channels and the constant evolution of viruses are making botnet identification a more difficult task. With the speedy expansion of the Internet of Things (IoT), several network devices have fallen victim to botnet assaults, resulting in significant losses across various industries. Botnets are a major threat to network security, but there is hope that deep learning models will help spot botnet activity in network traffic data more effectively. This study proposes a modified attention mechanism network (MANet) for detecting IoT threats. RBM extracts features, and WSOA selects crucial ones. The MANet, CMSAMB-UNet, combines CSAM and MSAB for enhanced feature extraction and global modelling. This study uses Resnet50 as the foundation for CMSAMB-UNet, extracting multi-level characteristics. Local attention-based CSAM enhances features and removes redundancies. Experiments on the BoT-IoT dataset show its potential for improving IoT security.