Natural language processing (NLP) models based on deep neural networks (DNNs) are vulnerable to backdoor attacks. Existing backdoor defense methods have limited effectiveness and coverage scenarios. We propose a textual backdoor defense method based on deep feature classification. The method includes deep feature extraction and classifier construction. The method exploits the distinguishability of deep features of poisoned data and benign data. Backdoor defense is implemented in both offline and online scenarios. We conducted defense experiments on two datasets and two models for a variety of backdoor attacks. The experimental results demonstrate the effectiveness of this defense approach and outperform the baseline defense method.

Deep learning models are vulnerable to backdoor attacks. The success rate of textual backdoor attacks based on data poisoning in existing research is as high as 100%. In order to enhance the natural language processing model’s defense against backdoor attacks, we propose a textual backdoor defense method via poisoned sample recognition. Our method consists of two parts: the first step is to add a controlled noise layer after the model embedding layer, and to train a preliminary model with incomplete or no backdoor embedding, which reduces the effectiveness of poisoned samples. Then, we use the model to initially identify the poisoned samples in the training set so as to narrow the search range of the poisoned samples. The second step uses all the training data to train an infection model embedded in the backdoor, which is used to reclassify the samples selected in the first step, and finally identify the poisoned samples. Through detailed experiments, we have proved that our defense method can effectively defend against a variety of backdoor attacks (character-level, word-level and sentence-level backdoor attacks), and the experimental effect is better than the baseline method. For the BERT model trained by the IMDB dataset, this method can even reduce the success rate of word-level backdoor attacks to 0%.

Recent studies have shown that natural language processing (NLP) models are vulnerable to adversarial examples, which are maliciously designed by adding small perturbations to benign inputs that are imperceptible to the human eye, leading to false predictions by the target model. Compared to character- and sentence-level textual adversarial attacks, word-level attack can generate higher-quality adversarial examples, especially in a black-box setting. However, existing attack methods usually require a huge number of queries to successfully deceive the target model, which is costly in a real adversarial scenario. Hence, finding appropriate models is difficult. Therefore, we propose a novel attack method, the main idea of which is to fully utilize the adversarial examples generated by the local model and transfer part of the attack to the local model to complete ahead of time, thereby reducing costs related to attacking the target model. Extensive experiments conducted on three public benchmarks show that our attack method can not only improve the success rate but also reduce the cost, while outperforming the baselines by a significant margin.

Despite deep neural networks (DNNs) having achieved impressive performance in various domains, it has been revealed that DNNs are vulnerable in the face of adversarial examples, which are maliciously crafted by adding human-imperceptible perturbations to an original sample to cause the wrong output by the DNNs. Encouraged by numerous researches on adversarial examples for computer vision, there has been growing interest in designing adversarial attacks for Natural Language Processing (NLP) tasks. However, the adversarial attacking for NLP is challenging because text is discrete data and a small perturbation can bring a notable shift to the original input. In this paper, we propose a novel method, based on conditional BERT sampling with multiple standards, for generating universal adversarial perturbations: input-agnostic of words that can be concatenated to any input in order to produce a specific prediction. Our universal adversarial attack can create an appearance closer to natural phrases and yet fool sentiment classifiers when added to benign inputs. Based on automatic detection metrics and human evaluations, the adversarial attack we developed dramatically reduces the accuracy of the model on classification tasks, and the trigger is less easily distinguished from natural text. Experimental results demonstrate that our method crafts more high-quality adversarial examples as compared to baseline methods. Further experiments show that our method has high transferability. Our goal is to prove that adversarial attacks are more difficult to detect than previously thought and enable appropriate defenses.