Agents
Backdoor attacks and defenses for semantic-communication systems over shared channels
This paper demonstrates backdoor attacks and defenses for semantic-communication (SemCom) systems operating over a multiple-access channel, showing how a poisoned encoder can be triggered to corrupt task-oriented transmission. It extends the model supply-chain and backdoor threat surface into learned wireless comms stacks. Off the core agent track, included as a low-importance security signal.
Source
↳ Follow the thread