This is the source code repository associated with the paper "Backdoors in Neural Models of Source Code", which is accepted for publication at ICPR'22. It was previously also presented at the NLP4Prog Workshop at ACL'21.

The arXiv version of the paper can be found here.

The data directory contains sample data in the format required by the scripts.
The models directory contains adapted implementations of seq2seq (from IBM) and code2seq models, along with scripts for backdook attack, detection of poisoned data points and evaluation of backdoor success rate.

The main script to run experiments is run.sh.