Attack traffic is the presence of malicious network traffic within a dataset, suggesting that at least one attack occurred. This type of traffic is typically generated by cybercriminals, hackers, or automated bots and can take many forms, including DDoS attacks, malware infections, and phishing efforts. To overcome this, a novel Attack Traffic detection (AT-DKNN) is proposed in this paper for effectively detecting and categorizing the types of attacks to improve cloud security. The process begins with data from the IoT device. Preprocessing of the collected dataset involves data normalization to standardize the inputs. The Spatio-Temporal Graph Neural Network (STGNN) is used for feature extraction, exploiting both spatial and temporal connections to create detailed feature representations. The feature extraction was followed by feature selection using Red Kite optimization to determine the most relevant attributes and reduce dimensionality. The optimized features are fed into a Deep Kronecker Neural Network, which classifies network data as normal or attack traffic. Attack traffic is recognized for further action, whereas normal traffic is safely routed to the cloud environment. The AT-DKSNN method achieves 98.33% accuracy, while the DNN, LSTM, and CNN approaches reach low accuracy of 96.26%, 92.32%, and 94.25%, respectively.