Jonathan Rusert

Jonathan Rusert, Zubair Shafiq, Padmini Srinivasan

Social media platforms are deploying machine learning based offensive language classification systems to combat hateful, racist, and other forms of offensive speech at scale. However, despite their real-world deployment, we do not yet comprehensively understand the extent to which offensive language classifiers are robust against adversarial attacks. Prior work in this space is limited to studying robustness of offensive language classifiers against primitive attacks such as misspellings and extraneous spaces. To address this gap, we systematically analyze the robustness of state-of-the-art offensive language classifiers against more crafty adversarial attacks that leverage greedy- and attention-based word selection and context-aware embeddings for word replacement. Our results on multiple datasets show that these crafty adversarial attacks can degrade the accuracy of offensive language classifiers by more than 50% while also being able to preserve the readability and meaning of the modified text.

Wanyue Zhai, Jonathan Rusert, Zubair Shafiq et al.

Recent advances in natural language processing have enabled powerful privacy-invasive authorship attribution. To counter authorship attribution, researchers have proposed a variety of rule-based and learning-based text obfuscation approaches. However, existing authorship obfuscation approaches do not consider the adversarial threat model. Specifically, they are not evaluated against adversarially trained authorship attributors that are aware of potential obfuscation. To fill this gap, we investigate the problem of adversarial authorship attribution for deobfuscation. We show that adversarially trained authorship attributors are able to degrade the effectiveness of existing obfuscators from 20-30% to 5-10%. We also evaluate the effectiveness of adversarial training when the attributor makes incorrect assumptions about whether and which obfuscator was used. While there is a a clear degradation in attribution accuracy, it is noteworthy that this degradation is still at or above the attribution accuracy of the attributor that is not adversarially trained at all. Our results underline the need for stronger obfuscation approaches that are resistant to deobfuscation

Jonathan Rusert, Padmini Srinivasan

Deep learning (DL) is being used extensively for text classification. However, researchers have demonstrated the vulnerability of such classifiers to adversarial attacks. Attackers modify the text in a way which misleads the classifier while keeping the original meaning close to intact. State-of-the-art (SOTA) attack algorithms follow the general principle of making minimal changes to the text so as to not jeopardize semantics. Taking advantage of this we propose a novel and intuitive defense strategy called Sample Shielding. It is attacker and classifier agnostic, does not require any reconfiguration of the classifier or external resources and is simple to implement. Essentially, we sample subsets of the input text, classify them and summarize these into a final decision. We shield three popular DL text classifiers with Sample Shielding, test their resilience against four SOTA attackers across three datasets in a realistic threat setting. Even when given the advantage of knowing about our shielding strategy the adversary's attack success rate is <=10% with only one exception and often < 5%. Additionally, Sample Shielding maintains near original accuracy when applied to original texts. Crucially, we show that the `make minimal changes' approach of SOTA attackers leads to critical vulnerabilities that can be defended against with an intuitive sampling strategy.

Osama Khalid, Jonathan Rusert, Padmini Srinivasan

Prior research has discussed and illustrated the need to consider linguistic norms at the community level when studying taboo (hateful/offensive/toxic etc.) language. However, a methodology for doing so, that is firmly founded on community language norms is still largely absent. This can lead both to biases in taboo text classification and limitations in our understanding of the causes of bias. We propose a method to study bias in taboo classification and annotation where a community perspective is front and center. This is accomplished by using special classifiers tuned for each community's language. In essence, these classifiers represent community level language norms. We use these to study bias and find, for example, biases are largest against African Americans (7/10 datasets and all 3 classifiers examined). In contrast to previous papers we also study other communities and find, for example, strong biases against South Asians. In a small scale user study we illustrate our key idea which is that common utterances, i.e., those with high alignment scores with a community (community classifier confidence scores) are unlikely to be regarded taboo. Annotators who are community members contradict taboo classification decisions and annotations in a majority of instances. This paper is a significant step toward reducing false positive taboo decisions that over time harm minority communities.

Jonathan Rusert

Text classification systems have continuously improved in performance over the years. However, nearly all current SOTA classifiers have a similar shortcoming, they process text in a horizontal manner. Vertically written words will not be recognized by a classifier. In contrast, humans are easily able to recognize and read words written both horizontally and vertically. Hence, a human adversary could write problematic words vertically and the meaning would still be preserved to other humans. We simulate such an attack, VertAttack. VertAttack identifies which words a classifier is reliant on and then rewrites those words vertically. We find that VertAttack is able to greatly drop the accuracy of 4 different transformer models on 5 datasets. For example, on the SST2 dataset, VertAttack is able to drop RoBERTa's accuracy from 94 to 13%. Furthermore, since VertAttack does not replace the word, meaning is easily preserved. We verify this via a human study and find that crowdworkers are able to correctly label 77% perturbed texts perturbed, compared to 81% of the original texts. We believe VertAttack offers a look into how humans might circumvent classifiers in the future and thus inspire a look into more robust algorithms.

Jonathan Rusert

In response to adversarial text attacks, attack detection models have been proposed and shown to successfully identify text modified by adversaries. Attack detection models can be leveraged to provide an additional check for NLP models and give signals for human input. However, the reliability of these models has not yet been thoroughly explored. Thus, we propose and test a novel attack setting and attack, RedHerring. RedHerring aims to make attack detection models unreliable by modifying a text to cause the detection model to predict an attack, while keeping the classifier correct. This creates a tension between the classifier and detector. If a human sees that the detector is giving an ``incorrect'' prediction, but the classifier a correct one, then the human will see the detector as unreliable. We test this novel threat model on 4 datasets against 3 detectors defending 4 classifiers. We find that RedHerring is able to drop detection accuracy between 20 - 71 points, while maintaining (or improving) classifier accuracy. As an initial defense, we propose a simple confidence check which requires no retraining of the classifier or detector and increases detection accuracy greatly. This novel threat model offers new insights into how adversaries may target detection models.