In recent years, Natural Language Generation (NLG) techniques have greatly advanced, especially in the realm of Large Language Models (LLMs). With respect to the quality of generated texts, it is no longer trivial to tell the difference between human-written and LLMgenerated texts (i.e., deepfake texts). While this is a celebratory feat for NLG, it poses new security risks (e.g., the generation of misinformation). To combat this novel challenge, researchers have developed diverse techniques to detect deepfake texts. While this niche field of deepfake text detection is growing, the field of NLG is growing at a much faster rate, thus making it difficult to understand the complex interplay between state-of-the-art NLG methods and the detectability of their generated texts. To understand such inter-play, two new computational problems emerge: (1) Deepfake Text Attribution (DTA) and (2) Deepfake Text Obfuscation (DTO) problems, where the DTA problem is concerned with attributing the authorship of a given text to one of k NLG methods, while the DTO problem is to evade the authorship of a given text by modifying parts of the text. In this cutting-edge tutorial, therefore, we call attention to the serious security risk both emerging problems pose and give a comprehensive review of recent literature on the detection and obfuscation of deepfake text authorships. Our tutorial will be 3 hours long with a mix of lecture and hands-on examples for interactive audience participation. You can find our tutorial materials here: https://tinyurl.com/naacl24-tutorial.

Consider a scenario where an author (e.g., activist, whistle-blower) with many public writings wishes to write “anonymously” when attackers may have already built an authorship attribution (AA) model based off of public writings including those of the author. To enable her wish, we ask a question “can one make the publicly released writings, T , unattributable so that AA models trained on T cannot attribute its authorship well?” Toward this question, we present a novel solution, UPTON, that exploits black-box data poisoning methods to weaken the authorship features in training samples and make released texts unlearnable. It is different from previous obfuscation works (e.g., adversarial attacks that modify test samples or backdoor works that only change the model outputs when triggering words occur). Using four authorship datasets (IMDb10, IMDb64, Enron and WJO), we present empirical validation where UPTON successfully downgrades the accuracy of AA models to the impractical level (e.g., ~ 35%) while keeping texts still readable (e.g., > 0.9 in BERTScore). UPTON remains effective to AA models that are already trained on available clean writings of authors.

Authorship Analysis, also known as stylometry, has been an essential aspect of Natural Language Processing (NLP) for a long time. Likewise, the recent advancement of Large Language Models (LLMs) has made authorship analysis increasingly crucial for distinguishing between human-written and AI-generated texts. However, these authorship analysis tasks have primarily been focused on written texts, not considering spoken texts. Thus, we introduce the largest benchmark for spoken texts - \sf HANSEN( ̲Human  ̲ANd ai  ̲Spoken t ̲Ext be ̲Nchmark). \sf HANSEN encompasses meticulous curation of existing speech datasets accompanied by transcripts, alongside the creation of novel AI-generated spoken text datasets. Together, it comprises 17 human datasets, and AI-generated spoken texts created using 3 prominent LLMs: ChatGPT, PaLM2, and Vicuna13B. To evaluate and demonstrate the utility of \sf HANSEN, we perform Authorship Attribution (AA) & Author Verification (AV) on human-spoken datasets and conducted Human vs. AI text detection using state-of-the-art (SOTA) models. While SOTA methods, such as, character n-gram or Transformer-based model, exhibit similar AA & AV performance in human-spoken datasets compared to written ones, there is much room for improvement in AI-generated spoken text detection. The \sf HANSEN benchmark is available at: https://huggingface.co/datasets/HANSEN-REPO/HANSEN

There is a lack of research into capabilities of recent LLMs to generate convincing text in languages other than English and into performance of detectors of machine-generated text in multilingual settings. This is also reflected in the available benchmarks which lack authentic texts in languages other than English and predominantly cover older generators. To fill this gap, we introduce MULTITuDE, a novel benchmarking dataset for multilingual machine-generated text detection comprising of 74,081 authentic and machine-generated texts in 11 languages (ar, ca, cs, de, en, es, nl, pt, ru, uk, and zh) generated by 8 multilingual LLMs. Using this benchmark, we compare the performance of zero-shot (statistical and black-box) and fine-tuned detectors. Considering the multilinguality, we evaluate 1) how these detectors generalize to unseen languages (linguistically similar as well as dissimilar) and unseen LLMs and 2) whether the detectors improve their performance when trained on multiple languages.

LIME has emerged as one of the most commonly referenced tools in explainable AI (XAI) frameworks that is integrated into critical machine learning applications (e.g., healthcare and finance). However, its stability remains little explored, especially in the context of text data, due to the unique text-space constraints. To address these challenges, in this paper, we first evaluate the inherent instability of LIME on text data to establish a baseline, and then propose a novel algorithm XAIFooler to perturb text inputs and manipulate explanations that casts investigation on the stability of LIME as a text perturbation optimization problem. XAIFooler conforms to the constraints to preserve text semantics and original prediction with small perturbations, and introduces Rank-biased Overlap (RBO) as a key part to guide the optimization of XAIFooler that satisfies all the requirements for explanation similarity measure. Extensive experiments on real-world text datasets demonstrate that XAIFooler significantly outperforms all baselines by large margins in its ability to manipulate LIME’s explanations with high semantic preservability.

Even though several methods have proposed to defend textual neural network (NN) models against black-box adversarial attacks, they often defend against a specific text perturbation strategy and/or require re-training the models from scratch. This leads to a lack of generalization in practice and redundant computation. In particular, the state-of-the-art transformer models (e.g., BERT, RoBERTa) require great time and computation resources. By borrowing an idea from software engineering, in order to address these limitations, we propose a novel algorithm, SHIELD, which modifies and re-trains only the last layer of a textual NN, and thus it “patches” and “transforms” the NN into a stochastic weighted ensemble of multi-expert prediction heads. Considering that most of current black-box attacks rely on iterative search mechanisms to optimize their adversarial perturbations, SHIELD confuses the attackers by automatically utilizing different weighted ensembles of predictors depending on the input. In other words, SHIELD breaks a fundamental assumption of the attack, which is a victim NN model remains constant during an attack. By conducting comprehensive experiments, we demonstrate that all of CNN, RNN, BERT, and RoBERTa-based textual NNs, once patched by SHIELD, exhibit a relative enhancement of 15%–70% in accuracy on average against 14 different black-box attacks, outperforming 6 defensive baselines across 3 public datasets. All codes are to be released.

The COVID-19 pandemic has created threats to global health control. Misinformation circulated on social media and news outlets has undermined public trust towards Government and health agencies. This problem is further exacerbated in developing countries or low-resource regions, where the news is not equipped with abundant English fact-checking information. In this paper, we make the first attempt to detect COVID-19 misinformation (in English, Spanish, and Haitian French) populated in the Caribbean regions, using the fact-checked claims in the US (in English). We started by collecting a dataset of Caribbean real & fake claims. Then we trained several classification and language models on COVID-19 in the high-resource language regions and transferred the knowledge to the Caribbean claim dataset. The experimental results of this paper reveal the limitations of current fake claim detection in low-resource regions and encourage further research on multi-lingual detection.

We proposes a novel algorithm, ANTHRO, that inductively extracts over 600K human-written text perturbations in the wild and leverages them for realistic adversarial attack. Unlike existing character-based attacks which often deductively hypothesize a set of manipulation strategies, our work is grounded on actual observations from real-world texts. We find that adversarial texts generated by ANTHRO achieve the best trade-off between (1) attack success rate, (2) semantic preservation of the original text, and (3) stealthiness–i.e. indistinguishable from human writings hence harder to be flagged as suspicious. Specifically, our attacks accomplished around 83% and 91% attack success rates on BERT and RoBERTa, respectively. Moreover, it outperformed the TextBugger baseline with an increase of 50% and 40% in terms of semantic preservation and stealthiness when evaluated by both layperson and professional human workers. ANTHRO can further enhance a BERT classifier’s performance in understanding different variations of human-written toxic texts via adversarial training when compared to the Perspective API.

Recent progress in generative language models has enabled machines to generate astonishingly realistic texts. While there are many legitimate applications of such models, there is also a rising need to distinguish machine-generated texts from human-written ones (e.g., fake news detection). However, to our best knowledge, there is currently no benchmark environment with datasets and tasks to systematically study the so-called ”Turing Test” problem for neural text generation methods. In this work, we present the TURINGBENCH benchmark environment, which is comprised of (1) a dataset with 200K human- or machine-generated samples across 20 labels Human, GPT-1, GPT-2_small, GPT-2_medium, GPT-2_large,GPT-2_xl, GPT-2_PyTorch, GPT-3, GROVER_base, GROVER_large, GROVER_mega, CTRL, XLM, XLNET_base, XLNET_large, FAIR_wmt19, FAIR_wmt20, TRANSFORMER_XL, PPLM_distil, PPLM_gpt2, (2) two benchmark tasks–i.e., Turing Test (TT) and Authorship Attribution (AA), and (3) a website with leaderboards. Our preliminary experimental results using TURINGBENCH show that GPT-3 and FAIR_wmt20 are the current winners, among all language models tested, in generating the most human-like indistinguishable texts with the lowest F1 score by five state-of-the-art TT detection models. The TURINGBENCH is available at: https://turingbench.ist.psu.edu/

The Universal Trigger (UniTrigger) is a recently-proposed powerful adversarial textual attack method. Utilizing a learning-based mechanism, UniTrigger generates a fixed phrase that, when added to any benign inputs, can drop the prediction accuracy of a textual neural network (NN) model to near zero on a target class. To defend against this attack that can cause significant harm, in this paper, we borrow the “honeypot” concept from the cybersecurity community and propose DARCY, a honeypot-based defense framework against UniTrigger. DARCY greedily searches and injects multiple trapdoors into an NN model to “bait and catch” potential attacks. Through comprehensive experiments across four public datasets, we show that DARCY detects UniTrigger’s adversarial attacks with up to 99% TPR and less than 2% FPR in most cases, while maintaining the prediction accuracy (in F1) for clean inputs within a 1% margin. We also demonstrate that DARCY with multiple trapdoors is also robust to a diverse set of attack scenarios with attackers’ varying levels of knowledge and skills. We release the source code of DARCY at: https://github.com/lethaiq/ACL2021-DARCY-HoneypotDefenseNLP.

In recent years, the task of generating realistic short and long texts have made tremendous advancements. In particular, several recently proposed neural network-based language models have demonstrated their astonishing capabilities to generate texts that are challenging to distinguish from human-written texts with the naked eye. Despite many benefits and utilities of such neural methods, in some applications, being able to tell the “author” of a text in question becomes critically important. In this work, in the context of this Turing Test, we investigate the so-called authorship attribution problem in three versions: (1) given two texts T1 and T2, are both generated by the same method or not? (2) is the given text T written by a human or machine? (3) given a text T and k candidate neural methods, can we single out the method (among k alternatives) that generated T? Against one humanwritten and eight machine-generated texts (i.e., CTRL, GPT, GPT2, GROVER, XLM, XLNET, PPLM, FAIR), we empirically experiment with the performance of various models in three problems. By and large, we find that most generators still generate texts significantly different from human-written ones, thereby making three problems easier to solve. However, the qualities of texts generated by GPT2, GROVER, and FAIR are better, often confusing machine classifiers in solving three problems. All codes and datasets of our experiments are available at: https://bit.ly/ 302zWdz