DAFF: Deployment-Aware Feature Fusion for Efficient TinyML-Based Intrusion Detection in IoT

Abstract

Tiny Machine Learning (TinyML) enables on-device inference for resource-constrained IoT systems, yet most intrusion detection approaches focus primarily on classification accuracy while overlooking deployment constraints such as latency, memory footprint, and energy consumption. This paper presents a system-level TinyML framework incorporating a novel Deployment-Aware Feature Fusion (DAFF) strategy, which adaptively combines statistical relevance and sparsitydriven importance to produce compact and discriminative feature subsets aligned with deployment requirements. The proposed framework evaluates multiple feature selection strategies, classical and neural models, and quantization schemes across six heterogeneous IoT intrusion datasets, while jointly analyzing system-level metrics such as inference latency, memory usage, and estimated energy consumption under microcontroller constraints. Experimental results show that DAFF preserves accuracy while reducing inference latency by up to 67%, lowering memory usage by up to 12%, and decreasing feature selection overhead by over 99% compared to LASSO, demonstrating the effectiveness of deployment-aware optimization for practical TinyML-based intrusion detection.