Fixed-Set Robustness in Programming by Example: Example Corruption and Semantic Partition Recovery

Short summary

This ArXiv paper investigates adversarial robustness in Programming-by-Example systems, where attackers can corrupt input-output examples to break program synthesis. It formalizes fixed-set worst-case corruption, introduces semantic partition aggregation as a defense mechanism, and demonstrates that low-margin PBE tasks are highly vulnerable to targeted attacks—though the proposed defense has limited effectiveness on realistic benchmarks and accepted public datasets.

• •Attackers can poison input-output examples to break PBE program synthesis, exploiting a gap standard noise-based evaluations miss
• •Semantic partition aggregation defense is proposed but proves ineffective on real-world benchmarks and accepted SyGuS slices
• •Low-margin PBE tasks are the main vulnerability vector; high-margin tasks remain robust

Generated with AI, which can make mistakes.