Backdoor attacks manipulate model predictions by inserting innocuous triggers into training and test data. We focus on more realistic and more challenging clean-label attacks where the adversarial training examples are correctly labeled. Our attack, LLMBkd, leverages language models to automatically insert diverse style-based triggers into texts. We also propose a poison selection technique to improve the effectiveness of both LLMBkd as well as existing textual backdoor attacks. Lastly, we describe REACT, a baseline defense to mitigate backdoor attacks via antidote training examples. Our evaluations demonstrate LLMBkd’s effectiveness and efficiency, where we consistently achieve high attack success rates across a wide range of styles with little effort and no model training.

Natural language processing (NLP) systems are often used for adversarial tasks such as detecting spam, abuse, hate speech, and fake news. Properly evaluating such systems requires dynamic evaluation that searches for weaknesses in the model, rather than a static test set. Prior work has evaluated such models on both manually and automatically generated examples, but both approaches have limitations: manually constructed examples are time-consuming to create and are limited by the imagination and intuition of the creators, while automatically constructed examples are often ungrammatical or labeled inconsistently. We propose to combine human and AI expertise in generating adversarial examples, benefiting from humans’ expertise in language and automated attacks’ ability to probe the target system more quickly and thoroughly. We present a system that facilitates attack construction, combining human judgment with automated attacks to create better attacks more efficiently. Preliminary results from our own experimentation suggest that human-AI hybrid attacks are more effective than either human-only or AI-only attacks. A complete user study to validate these hypotheses is still pending.

Adversarial attacks curated against NLP models are increasingly becoming practical threats. Although various methods have been developed to detect adversarial attacks, securing learning-based NLP systems in practice would require more than identifying and evading perturbed instances. To address these issues, we propose a new set of adversary identification tasks, Attacker Attribute Classification via Textual Analysis (AACTA), that attempts to obtain more detailed information about the attackers from adversarial texts. Specifically, given a piece of adversarial text, we hope to accomplish tasks such as localizing perturbed tokens, identifying the attacker’s access level to the target model, determining the evasion mechanism imposed, and specifying the perturbation type employed by the attacking algorithm. Our contributions are as follows: we formalize the task of classifying attacker attributes, and create a benchmark on various target models from sentiment classification and abuse detection domains. We show that signals from BERT models and target models can be used to train classifiers that reveal the properties of the attacking algorithms. We demonstrate that adversarial attacks leave interpretable traces in both feature spaces of pre-trained language models and target models, making AACTA a promising direction towards more trustworthy NLP systems.

Evaluating on adversarial examples has become a standard procedure to measure robustness of deep learning models. Due to the difficulty of creating white-box adversarial examples for discrete text input, most analyses of the robustness of NLP models have been done through black-box adversarial examples. We investigate adversarial examples for character-level neural machine translation (NMT), and contrast black-box adversaries with a novel white-box adversary, which employs differentiable string-edit operations to rank adversarial changes. We propose two novel types of attacks which aim to remove or change a word in a translation, rather than simply break the NMT. We demonstrate that white-box adversarial examples are significantly stronger than their black-box counterparts in different attack scenarios, which show more serious vulnerabilities than previously known. In addition, after performing adversarial training, which takes only 3 times longer than regular training, we can improve the model’s robustness significantly.

We propose an efficient method to generate white-box adversarial examples to trick a character-level neural classifier. We find that only a few manipulations are needed to greatly decrease the accuracy. Our method relies on an atomic flip operation, which swaps one token for another, based on the gradients of the one-hot input vectors. Due to efficiency of our method, we can perform adversarial training which makes the model more robust to attacks at test time. With the use of a few semantics-preserving constraints, we demonstrate that HotFlip can be adapted to attack a word-level classifier as well.

Classifying the stance expressed in online microblogging social media is an emerging problem in opinion mining. We propose a probabilistic approach to stance classification in tweets, which models stance, target of stance, and sentiment of tweet, jointly. Instead of simply conjoining the sentiment or target variables as extra variables to the feature space, we use a novel formulation to incorporate three-way interactions among sentiment-stance-input variables and three-way interactions among target-stance-input variables. The proposed specification intuitively aims to discriminate sentiment features from target features for stance classification. In addition, regularizing a single stance classifier, which handles all targets, acts as a soft weight-sharing among them. We demonstrate that discriminative training of this model achieves the state-of-the-art results in supervised stance classification, and its generative training obtains competitive results in the weakly supervised setting.