Anti-Backdoor Learning: Training Clean Models on Poisoned Data

Backdoor attack has emerged as a major security threat to deep neural networks (DNNs). While existing defense methods have demonstrated promising results on detecting and erasing backdoor triggers, it is still not clear if measures can be taken to avoid the triggers from being learned into the model in the first place. In this paper, we introduce the concept of \emph{anti-backdoor learning}, of which the aim is to train clean models on backdoor-poisoned data. We frame the overall learning process as a dual-task of learning the clean portion of data and learning the backdoor portion of data. From this view, we identify two inherent characteristics of backdoor attacks as their weaknesses: 1) the models learn backdoored data at a much faster rate than learning clean data, and the stronger the attack the faster the model converges on backdoored data; and 2) the backdoor task is tied to a specific class (the backdoor target class). Based on these two weaknesses, we propose a general learning scheme, Anti-Backdoor Learning (ABL), to automatically prevent backdoor attacks during training. ABL introduces a two-stage \emph{gradient ascent} mechanism into standard training to 1) help isolate backdoor examples at an early training stage, and 2) break the correlation between backdoor examples and the target class at a later training stage. Through extensive experiments on multiple benchmark datasets against 10 state-of-the-art attacks, we empirically show that ABL-trained models on backdoor-poisoned data achieve the same performance as they were trained on purely clean data. Code is available at \underline{https://github.com/bboylyg/ABL}.