Rajvardhan Oak, Muhammad Haroon, Claire Wonjeong Jo, Magdalena Wojcieszak, Anshuman Chhabra. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Customer reviews play an important role in rankings and visibility on e-commerce sites, and also strongly influence a customer's decision to purchase a product. Motivated by this, malicious sellers engage in incentivized review fraud to inflate their product ratings by providing customers with free products in exchange for five-star reviews, thus compromising review integrity. While there is ample prior work on fake reviews in general, there is limited prior work on incentivized review fraud. In this work, we infiltrate an underground market for fake reviews and implement a custom crawler to collect a dataset of malicious products that seek incentivized reviews. We devise and extract a set of features, and show that these are statistically significant in differentiating between benign and malicious products. Using hypothesis testing, we identify characteristics and trends exhibited by malicious products. While we are unable to achieve a high precision when we train standard machine learning models without compromising on the recall, we propose a lightweight two-phase technique that combines high-precision product classifiers with high-recall review classifiers. This technique allows us to minimize the false positives, with only a slight increase in false negatives. Finally, we also audit the effectiveness of two publicly available tools for incentivized review detection and find that they are not reliable. In summary, we contribute a new high-fidelity dataset, characterize products seeking incentivized reviews, audit existing tools for review analysis, and present a superior method for detecting review fraud. We hope that this research could be useful for e-commerce companies and other entities who have a stake in preserving opinion and review integrity online.

While human factors in fraud have been studied by the HCI and security communities, most research has been directed to understanding either the victims’ perspectives or prevention strategies, and not on fraudsters, their motivations and operation techniques. Additionally, the focus has been on a narrow set of problems: phishing, spam and bullying. In this work, we seek to understand review fraud on e-commerce platforms through an HCI lens. Through surveys with real fraudsters (N=36 agents and N=38 reviewers), we uncover sophisticated recruitment, execution, and reporting mechanisms fraudsters use to scale their operation while resisting takedown attempts, including the use of AI tools like ChatGPT. We find that countermeasures that crack down on communication channels through which these services operate are effective in combating incentivized reviews. This research sheds light on the complex landscape of incentivized reviews, providing insights into the mechanics of underground services and their resilience to removal efforts.

Authorship obfuscation is the process of making changes to text such that identifying attributes (style, common words and phrases, tone) are masked. The goal of obfuscation is to retain the semantics of the text (i.e., the meaning) but rewrite it in such a way that the author cannot be identified. In this work, we investigate the effectiveness of language models for authorship obfuscation. More specifically, we examine the application of document summarization (a task where we learn to generate the summary of a text) as an authorship obfuscation method. Since summaries are shorter versions of text but which retain the significant points made in it, we hypothesize that summaries will be stripped off any stylistic identifying features of the text. Our experiments show that this is indeed the case; we were able to fool authorship classifiers and degrade their performance by as much as 70% However, this also significantly affected the semantics; there was a non-trivial loss of information and the produced text was not an accurate representation of the original.

Anomaly detection is essential towards ensuring system security and reliability. Powered by constantly generated system data, deep learning has been found both effective and flexible to use, with its ability to extract patterns without much domain knowledge. Existing anomaly detection research focuses on a scenario referred to as zero-positive, which means that the detection model is only trained for normal (i.e., negative) data. In a real application scenario, there may be additional manually inspected positive data provided after the system is deployed. We refer to this scenario as lifelong anomaly detection. However, we find that existing approaches are not easy to adopt such new knowledge to improve system performance. In this work, we are the first to explore the lifelong anomaly detection problem, and propose novel approaches to handle corresponding challenges. In particular, we propose a framework called unlearning, which can effectively correct the model when a false negative (or a false positive) is labeled. To this aim, we develop several novel techniques to tackle two challenges referred to as exploding loss and catastrophic forgetting. In addition, we abstract a theoretical framework based on generative models. Under this framework, our unlearning approach can be presented in a generic way to be applied to most zero-positive deep learning-based anomaly detection algorithms to turn them into corresponding lifelong anomaly detection solutions. We evaluate our approach using two state-of-the-art zero-positive deep learning anomaly detection architectures and three real-world tasks. The results show that the proposed approach is able to significantly reduce the number of false positives and false negatives through unlearning.

With the advent of the Internet, social media platforms have become an increasingly popular medium of communication for people. Platforms like Twitter and Quora allow people to express their opinions on a large scale. These platforms are, however, plagued by the problem of hate speech and toxic content. Such content is generally sexist, homophobic or racist. Automatic text classification can filter out toxic content so some extent. In this paper, we discuss the adversarial attacks on hate speech classifiers. We demonstrate that by changing the text slightly, a classifier can be fooled to misclassifying a toxic comment as acceptable. We attack hate speech classifiers with known attacks as well as introduce four new attacks. We find that our method can degrade the performance of a Random Forest classifier by 20%. We hope that our work sheds light on the vulnerabilities of text classifiers, and opens doors for further research on this topic.

Generation of secure random numbers has always been a challenging issue in design and development of secure computer systems. Random numbers have important applications in the field of cryptography where the security of the scheme relies upon the random nature of the keys. It is not practically possible to achieve true randomness in a machine, and hence we rely upon Pseudo Random Number Generators (PRNGs) to produce near-true randomness. PRNGs use a mathematical function that relies upon a seed (a preset value required by the function to generate values) and it generates numbers which satisfy certain tests for randomness and appear to be random for a user having no knowledge of the generator function. These pseudorandom functions have their drawbacks due to them being derived from a mathematical function. To generate random numbers that can never be predicted by any observer, requires a causally non-deterministic process where events are not fully determined by prior states. Due to the physical impossibility of acquiring sufficient information to predict the outcome of such an event, its outcomes are guaranteed to be random to all. Various methods to generate pseudorandomness have been employed over the years which includes using mathematical functions, keyboard typing latency of the user, network latency, memory latency etc. as sources of generating random numbers. In this work, we propose a new way of generating pseudorandom numbers using generative adversarial networks. We demonstrate that a GAN can act as a Cryptographically Secure Pseudorandom Number Generator (CPRNG) passing 97% of National Institute of Standards and Technology (NIST) tests.