Browse Teach Curate

Adaptive Model Ensembling Defense Against Variable-Power Semantic Jamming Attacks

A defense for deep-learning semantic communication that first detects whether an adversarial semantic-jamming signal is present on the wireless channel and estimates its power, then selects a correspondingly robust member of a model ensemble, deliberately trading generalization for robustness only when strong jamming is detected. Paired with an adjustable-power perturbation generator as the threat model, it addresses the fixed-power limitation of prior semantic-jamming attacks and defenses.

Demerzel Text

ROME: Robust Model Ensembling for Semantic Communication Against Semantic Jamming Attacks Kequan

Deep-learning semantic communication systems transmit task-relevant meaning extracted by neural encoders instead of raw bits, which leaves them open to semantic jamming: crafted adversarial perturbat…