BCIwiki (bciwiki.com) — Researchers at the eBRAIN Lab, New York University Abu Dhabi, UAE, introduced ORBITAL, combining EEG decoder compression, adversarial robustness evaluation and edge-hardware tests. The manuscript was posted to arXiv on October 26, 2026; the authors report submission to IEEE ISCAS 2027 for review.
The study addresses compact electroencephalography (EEG) decoders for embedded brain-computer interfaces (BCIs). ORBITAL combines hardware-realizable structured compression, recovery-aware operating-point selection, multi-attack evaluation, and native edge validation under robustness constraints. Adversarial robustness here concerns maintaining correct predictions under the evaluated attacks, rather than resistance to every source of noise or real-world risk.
Across nine subjects and three random seeds on BCI Competition IV-2a, S25-TRADES achieved the highest observed mean true-union robustness among the tested compressed EEGNet models: 23.06%, versus 21.27% for standard Full. It used 26.27% fewer parameters and 25.10% fewer multiply-accumulate operations (MACs) than standard Full. In matched controls, TRADES improved S25 union accuracy by 6.74 percentage points over clean recovery training (p=0.0039). The authors identify TRADES as the main source of robustness recovery, while their per-example analysis indicates decision-boundary redistribution rather than uniform expansion of classification margins.
On Jetson Orin Nano, FP16 compression reduced board energy by up to 16.20% and latency by 8.00%. Despite 99.3% prediction agreement, however, native INT8 quantization/dequantization (Q/DQ) incurred about twice the FP16 latency and energy. The authors interpret this as evidence that compression and low-precision deployment require joint consideration of adversarial robustness and hardware performance, rather than relying on model size or numerical precision alone.