IMBERT: Making BERT Immune to Insertion-based Backdoor Attacks

TL;DR

IMBERT is a novel method that leverages gradients or self-attention scores to detect and defend against insertion-based backdoor attacks in NLP models, significantly reducing attack success while maintaining accuracy.

Contribution

IMBERT introduces a self-defense mechanism for NLP models against insertion-based backdoor attacks using gradients and self-attention scores, with high detection accuracy and model-agnostic applicability.

Findings

Detects up to 98.5% of inserted triggers

Reduces attack success rate significantly

Maintains competitive accuracy on clean data

Abstract

Backdoor attacks are an insidious security threat against machine learning models. Adversaries can manipulate the predictions of compromised models by inserting triggers into the training phase. Various backdoor attacks have been devised which can achieve nearly perfect attack success without affecting model predictions for clean inputs. Means of mitigating such vulnerabilities are underdeveloped, especially in natural language processing. To fill this gap, we introduce IMBERT, which uses either gradients or self-attention scores derived from victim models to self-defend against backdoor attacks at inference time. Our empirical studies demonstrate that IMBERT can effectively identify up to 98.5% of inserted triggers. Thus, it significantly reduces the attack success rate while attaining competitive accuracy on the clean dataset across widespread insertion-based attacks compared to two baselines. Finally, we show that our approach is model-agnostic, and can be easily ported to several pre-trained transformer models.