The proliferation of Internet of Things devices in security-sensitive applications creates a need for lightweight anomaly detection mechanisms that can operate on resource-constrained embedded platforms. This paper presents asymmetric autoencoder architectures specifically designed for security monitoring on embedded devices. The proposed architectures employ different encoder and decoder complexities, with lightweight encoders suitable for resource-constrained devices and more complex decoders deployed on edge or cloud infrastructure. We evaluate several architecture variants including standard asymmetric autoencoders, variational asymmetric autoencoders, and sparse asymmetric autoencoders across multiple IoT security datasets. Experimental results demonstrate that the asymmetric approach reduces memory requirements by 62% and computational latency by 47% compared to symmetric architectures while maintaining detection accuracy within 2.1% of symmetric baselines. The sparse asymmetric autoencoder variant achieves the best overall performance, with 94.7% detection accuracy and only 340 KB memory footprint.
Mohamed et al. (Tue,) studied this question.