MariIRQ-Mamba-Lite: Lightweight CPU Interrupt Side-Channel Fingerprinting for Maritime Websites

Website visits to maritime digital services may be exposed through interrupt-related side channels arising during browser execution. To address the trade-off between recognition accuracy and computational cost in website fingerprinting, we propose MariIRQ-Adapt and the lightweight MariIRQ-Mamba-Lite. Adapt represents trace variations using an interrupt-aware weighted time counting matrix (IRQ-WTCM) and performs classification with a state-space model. Lite introduces a windowed interrupt counting matrix (WICM) combined with bidirectional Mamba2 modeling, reducing the feature elements per trace from 5000 to 1000, a reduction of 80%. On a maritime-related case-study dataset containing 100 public websites, five-seed Linux closed-world experiments yield accuracies of 95.76±0.59% for Adapt and 93.64±0.90% for Lite. Lite contains only 2.28 million parameters and achieves a mean per-seed 50th-percentile (p50) processing latency of approximately 24.61 ms from an already acquired trace to its predicted label on the evaluated graphics processing unit (GPU). Comparisons with multilayer perceptron (MLP), one-dimensional convolutional neural network (1D-CNN), convolutional neural network–long short-term memory (CNN–LSTM), and Transformer baselines, together with closed-world, temporal-drift, and background-workload evaluations on Linux-, Windows-, and macOS-acquired data, characterize recognition performance and its applicability. The results demonstrate effective website identification with compact representations, while showing that temporal drift and concurrent activity remain important sources of performance degradation. This study provides experimental evidence for assessing side-channel privacy risks in maritime-related browsing.