An AI-Enhanced Lightweight Data Integrity Framework for Secure IoT Communication

Abstract

When billions of everyday objects connect to open networks, two security problems tend to arise in parallel: attackers who alter sensor readings mid-transit, and attackers who masquerade as legitimate devices to inject fabricated data. This paper presents the AI-Enhanced Data Integrity Service (AI-DIS), a security framework tailored for the kinds of low-power IoT microcontrollers that standard protocols simply cannot accommodate. Building on the classic Data Integrity Service model—with its Source-Side Entity (SE) and Destination-Side Entity (DE)—AI-DIS introduces three coordinated layers of protection: a TinyML anomaly pre-filter that inspects outgoing sensor data at the SE before any signing takes place; a contextual behavioural validator at the DE that reasons about plausibility once the cryptographic check has passed; and a Federated Averaging (FedAvg) update loop that continuously refines the shared threat model without pulling raw sensor data away from the device. Each packet is signed using HMAC-SHA256 over the device MAC address, the smart-environment identifier, a strictly increasing sequence counter, and a timestamp—a design choice that rules out replay attacks at the protocol level rather than relying on detection heuristics. Testing on the N-BaIoT and IoT-FADS benchmark datasets with 5-fold cross-validation shows the LSTM-based pre-filter reaching F1 = 0.961 ± 0.007, a false-positive rate of only 0.9%, and an inference time of 12.1 ± 0.8 ms on an ARM Cortex-M4 running at 80 MHz. End-to-end, AI-DIS consumes 25.4 µJ per packet. A smart-home deployment study and a comparison spanning ten related works—two of them from 2026—show that AI-DIS is the only solution that satisfies all five required properties at once: data integrity, source authentication, lightweight operation, IoT-readiness, and AI-driven threat detection.