Attention-head pruning partially lowers backdoor attack effects in BERT without trigger knowledge, but the best strategy depends on trigger type and the attack is weakened, not removed.
In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Pruning Strategies for Backdoor Defense in LLMs
Attention-head pruning partially lowers backdoor attack effects in BERT without trigger knowledge, but the best strategy depends on trigger type and the attack is weakened, not removed.