IMBERT: Making BERT Immune to Insertion-based Backdoor Attacks

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. * Now at Google DeepMind. 1 According to statistics from Hugging Face, BERT receives 15M downloads per month.

Security Challenges in Natural Language Processing Models

“…The former method aims at identifying poisoned data by analyzing the anomalous characteristics of the training data He et al, 2023b). The latter approach leverages external tools (Qi et al, 2021a) or the victim language models themselves (Yang et al, 2021b;He et al, 2023a) to either remove the triggers or entirely discard the poisoned data samples during the inference.…”