IoT Network Intrusion Detection with Deep Learning and Voice Alerts

The growing complexity and scale of Internet of Things (IoT) networks have made them a prime target for cyber-attacks, necessitating the creation of Intrusion Detection Systems (IDS) to secure confidential data. A key challenge in existing IDS models is that a high loss value can indicate inadequate detection capabilities, potentially leading to missed attacks especially in time-sensitive environments. This study proposes a robust IDS framework by integrating an Autoencoder for feature extraction with Temporal Fusion Transformer (TFT) for attack detection of network intrusion in IoT environments. UNSW-NB15 dataset, was used to assess the model's efficiency. Data preprocessing techniques, including data cleaning and the applying the SMOTE, were employed to ensure balanced dataset representation. The Autoencoder's encoding layer reduces dimensionality to capture essential features and extracted that feature and that as an input to models like the TFT. The Temporal Fusion Transformer (TFT) leverages LSTM encoding, temporal attention, and Gated Residual Networks to capture complex time-series dependencies with the extracted features which help to detect attacks like Fuzzers, Dos, Backdoor, Exploits, Generic, Reconnaissance, Shellcode, Worms. After detecting the attacks, the voice alert is automatically sent to the admin which helps to maintain the robustness. The accuracy achieved by the proposed IDS is 99.06%, with a recall of 98.66%, precision of 91.59%, F1-score of 94.27% and a loss of 0.030 when compared to other traditional algorithms. These results demonstrate the model's effectiveness in identifying various types of cyber-attacks while maintaining a low loss when compared to others. The integration of the voice alert system further enhances the security of the IDS by providing immediate notifications of detected intrusion, thereby facilitating timely responses.