Homomorphic-Encrypted Inference
Computation directly on encrypted data using homomorphic encryption, so that a decoder could in principle run on ciphertext neural features and return an encrypted result that only the user can decrypt, without any server ever seeing the plaintext. Related secure-computation tools — secure multi-party computation and functional encryption — target the same goal by different means.
For BCI this remains largely aspirational: practical schemes (such as CKKS for approximate arithmetic) impose orders-of-magnitude overhead in compute and latency, which conflicts with the tens-of-milliseconds closed-loop budgets and the power limits of neural systems. It is most plausible today for offline or batch neural analytics and for outsourcing heavy computation, rather than for real-time on-implant decoding.
Encryption protects the confidentiality of a computation, not the correctness of the model or the privacy of its outputs; it composes with, rather than replaces, differential privacy and integrity checks.