Skip to content
InfoResearchPeer-reviewed

Compile-Time Backdoor Neutralization via LLM-Guided Localized Edits

Published
Record updated
View JSON

Summary

CBN is a defense framework against compile-time backdoors in deep learning compilers such as TVM, where a trigger effect is re-materialized during compilation. It performs localized rewriting at the compiler-IR level, guided by runtime probes and LLM-proposed plans from a finite edit vocabulary. Across seven ONNX-Zoo models and five commercial LLMs, it reduces mean DCL-BD attack success from 30.1% to 1.9% with 2.0% mean inference-latency overhead.

Mitigation

CBN: a defense framework for compilation-induced backdoors that performs localized rewriting at the compiler-IR level, guided by runtime probes and LLM-proposed plans drawn from a finite edit vocabulary, with a legality-aware schema, a budgeted-transformation validator, and a probe-based feedback loop that repairs broken plans.