<![CDATA[Issues in securing the resilience of Deep Learning models and operational problems related to model size and latency in the Internet of Things (IoT) environment, especially when applied to devices with limited resources, can be overcome by implementing a modified Quantization-Aware Training (QAT) model based on IDS and an autoencoder. This research proposes the use of QAT to achieve operational efficiency while strengthening the detection model’s resilience against data manipulation in MiTM attacks. By integrating QAT, the model not only becomes lighter for edge devices but also more capable of maintaining detection integrity even when model weights are compromised. The research was conducted using the CICIoT2023 dataset, and the output of the advanced QAT model was then applied to resource-constrained IoT devices, namely the Raspberry Pi 3. The methodology began with data collection, followed by data preprocessing, which was then normalized before being fed into the IDS and autoencoder techniques. After the autoencoder model was successfully created, the QAT model was developed so that it remained unchanged or even improved when implemented on the Raspberry Pi 3. With the application of QAT modifications, an 80.43% reduction in model size and an 11.2% increase in model inference speed were confirmed. Furthermore, when faced with QAT model weight corruption of up to 20%, the F1-score value remained stable at 100%. These results show that the QAT model is highly effective at improving model reliability and maintaining the resilience of deep learning models against MiTM attacks, even under limited conditions.]]>
