This work focuses on defending against the data poisoning based backdoor attacks, which bring in serious security threats to deep neural networks (DNNs). Specifically, given a untrustworthy training dataset, we aim to filter out potential poisoned samples, \ie, poisoned sample detection (PSD). The key solution for this task is to find a discriminative metric between clean and poisoned samples, even though there is no information about the potential poisoned samples (\eg, the attack method, the poisoning ratio). In this work, we develop an innovative detection approach from the perspective of the gradient \wrt activation (\ie, activation gradient direction, AGD) of each sample in the backdoored model trained on the untrustworthy dataset. We present an interesting observation that the circular distribution of AGDs among all samples of the target class is much more dispersed than that of one clean class. Motivated by this observation, we firstly design a novel metric called Cosine similarity Variation towards Basis Transition (CVBT) to measure the circular distribution's dispersion of each class. Then, we design a simple yet effective algorithm with identifying the target class(es) using outlier detection on CVBT scores of all classes, followed by progressively filtering of poisoned samples according to the cosine similarities of AGDs between every potential sample and a few additional clean samples. Extensive experiments under various settings verify that given very few clean samples of each class, the proposed method could filter out most poisoned samples, while avoiding filtering out clean samples, verifying its effectiveness on the PSD task. Codes are available at https://github.com/SCLBD/bdzoo2/blob/dev/detection_pretrain/agpd.py.

翻译：暂无翻译