Dynamic Semantic Equivalence Checking for Adversarially Robust Neural Compilers

ICLR 2026 Conference Submission25461 Authors

Keywords: Robust Neural Compilers

Abstract: We propose a novel architecture of neural compiler which incorporates a Dynamic Semantic Equivalence Checker (DSEC) to overcome the music of adversarial robustness in program compilation. Traditional neural compiler In both cases, traditional neural compilers are susceptible to the adversarial perturbations of input program code, leading to semantically incorrect optimization; the DSEC addresses this issue by using runtime verification when combined with probabilistic program analysis. The core innovation is the Relational Execution Tracker, a dynamic technique to compare execution traces of the original and compiled programs with the aid of a probabilistic divergence metric to identify behavioral differences. Furthermore, a Bayesian neural network-based Probabilistic Program Analyzer is used to assign perturbation likelihood estimate which allows targeted trace comparisons and efficient resource allocation. The system adaptively changes the optimization strategies in case of detection of adversarial influence, relying on a formally verified code generator for critical code regions. We also propose a hybrid adversarial training objective which constrains semantic consistency in addition to standard compilation accuracy.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 25461