Evidence · DiggingBeagle record
Data-Efficient Backdoor Attacks on Code Models via Trigger-Neutralized Counterfactual Scoring
Data-Efficient Backdoor Attacks on Code Models via Trigger-Neutralized Counterfactual Scoring
- Source role
- primary disclosure
Evidence record
Data-Efficient Backdoor Attacks on Code Models via Trigger-Neutralized Counterfactual Scoring
Claim-level citations (2)
- supportsTraining-data poison selection implants code-model backdoors at low sample rates: TNCS selects training poison candidates by contrasting scores for trigger-bearing inputs with trigger-neutralized counterparts, addressing both classification and autoregressive code generation.
Abstract, trigger-neutralized counterfactual scoring and model tasks
- supportsTraining-data poison selection implants code-model backdoors at low sample rates: The researchers compare triggered and trigger-neutralized sample scores to select high-effect training poisons and report attack success gains over several baselines at low poisoning ratios for three code-understanding and generation tasks.
Abstract: trigger-neutralized counterfactual scoring; three CodeXGLUE tasks; five trigger families; low-ratio study