ARTFEED — Contemporary Art Intelligence

SemBugger: Polymorphic Backdoor Attack on Semantic Communication

ai-technology · 2026-04-29

Researchers propose SemBugger, a polymorphic backdoor attack targeting Semantic Communication (SC) systems. Unlike monomorphic backdoors with a single attack target, SemBugger uses graded-intensity triggers to dynamically control SC knowledge, enabling diverse malicious outputs. The attack employs a multi-effect poisoning-training framework with hierarchical malicious loss, preserving transmission fidelity while manipulating system outputs. This approach enhances attack diversity, efficiency, and flexibility for heterogeneous downstream scenarios.

Key facts

  • SemBugger is a polymorphic backdoor attack for Semantic Communication systems.
  • It uses graded-intensity triggers to dynamically control SC knowledge.
  • The attack framework includes multi-effect poisoning-training and hierarchical malicious loss.
  • It overcomes limitations of monomorphic backdoors with single attack targets.
  • SemBugger preserves transmission fidelity while generating diverse malicious outputs.
  • The approach improves attack diversity, efficiency, and flexibility.
  • It targets heterogeneous downstream scenarios.
  • The research is published on arXiv (2604.23231).

Entities

Institutions

  • arXiv

Sources