Data Poisoning & Backdoor Attacks

Attacks that corrupt a decoder during learning rather than at inference. In poisoning, an adversary contaminates training or calibration data so the model's overall accuracy degrades or a specific behavior is implanted; in a backdoor, a hidden trigger pattern causes a chosen output only when it is present, while normal performance masks the tampering.

Adaptive and self-recalibrating BCIs are a natural target, because they continually incorporate new, often unlabeled data — the very mechanism that keeps them accurate through drift also opens a channel to steer them. Federated training widens the surface further, since malicious participants can submit poisoned updates. Robust aggregation, provenance checks on calibration data, and anomaly detection are the main countermeasures.