ORBITAL: Optimizing Robust Brain-Computer Inference via Architecture-Aware Low-Precision Compression

Researchers propose the ORBITAL framework, which uses architecture-aware low-precision compression to give brain-computer interface EEG decoders both adversarial robustness and hardware efficiency. On the nine subjects of BCI Competition IV-2a, S25-TRADES achieves the highest average true-union robustness among compressed EEGNet models at 23.06% (standard Full is 21.27%), with parameters and MACs reduced by 26.27% and 25.10% respectively. On Jetson Orin Nano, FP16 compression reduces board-level energy consumption by up to 16.20% and latency by 8.00%, while native INT8 Q/DQ, despite a 99.3% prediction agreement rate, has roughly twice the latency and energy consumption of FP16.