The Enemy in Your Own Camp: How Well Can We Detect Statistically-Generated Fake Reviews – An Adversarial Study

By Dirk Hovy. 2016

56 citations recorded by the source corpus.

Online reviews are a growing market, but it is struggling with fake reviews. They undermine both the value of reviews to the user, and their trust in the review sites. However, fake positive reviews can boost a business, and so a small industry producing fake reviews has developed. The two sides are facing an arms race that involves more and more natural language processing (NLP). So far, NLP has been used mostly for detection, and works well on human-generated reviews. But what happens if NLP techniques are used to generate fake reviews as well? We investigate the question in an adversarial setup, by assessing the detectability of different fake-review generation strategies. We use generative models to produce reviews based on meta-information, and evaluate their effectiveness against deceptiondetection models and human judges. We find that meta-information helps detection, but that NLP-generated reviews conditioned on such information are also much harder to detect than conventional ones.

Subfields: Information Systems, Sociology and Political Science, Safety, Robustness & Interpretability, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Information Systems

In Sociology and Political Science

In Safety, Robustness & Interpretability

In Signal Processing

Citations in the corpus

GLTR: Statistical Detection and Visualization of Generated Text (2019, background): “Hovy (2016) ﬁnds that a logistic regression model can detect generated product reviews at a higher rate than human judges, indicating that humans struggle with this task.”
Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense (2023, methodology): “Early efforts in this vein use classifiers to detect fake reviews (Hovy, 2016) and fake news (Zellers et al.”
Keep Calm and Switch On! Preserving Sentiment and Fluency in Semantic Text Exchange (2019, background): “Review generation (Hovy, 2016; Lipton et al., 2015; Dong et al., 2017; Yao et al., 2017; Juuti et al., 2018) comes with many forms.”
The Importance of Modeling Social Factors of Language: Theory and Practice (2021, background): “Consequently, NLP systems struggle with applications related to pragmatics and interaction, or when “what is said is not what is meant,” e.g., sarcasm, irony, deception, and any other situation that requires a “social” interpretation (Aber-crombie and Hovy, 2016).”
Automatic Detection of Machine Generated Text: A Critical Survey (2020, background): “Such auxiliary signals can be complementary to linguistic signals from the textual source for the detection task (Hovy, 2016; Solaiman et al., 2019).”
Provable Robust Watermarking for AI-Generated Text (2023, unknown): “One camp, inspired by Turing (1950), aims at generically distinguishing machine-generated text from that of the humans (Gehrmann et al., 2019; Mitchell et al., 2023; Hovy, 2016; Zellers et al., 2019; OpenAI, 2023b).”
Automated Crowdturfing Attacks and Defenses in Online Review Systems (2017, background): “Hovy [17] performed a preliminary investigation of n-gram-based review generation in an adversarial setting.”