Wednesday 26 March 2025
In recent years, the rapid development of large language models has enabled machines to generate highly human-like texts. While this technology holds great promise for various applications, it also raises concerns about the unrestricted dissemination of non-attributed textual contents, including misinformation and fabricated news.
To address these issues, a team of researchers has proposed a novel framework for generating adversarial examples in machine-generated text detection. The goal is to develop a robust method that can effectively defend against malicious perturbations and attacks.
The framework, dubbed GREATER-A, consists of two key components: an adversary and a detector. The adversary, also known as GREATER-Adversary, identifies critical tokens in the embedding space and perturbs them using greedy search and pruning to generate stealthy and disruptive adversarial examples. On the other hand, the detector, referred to as GREATER-Detector, learns to defend against these attacks by updating its parameters synchronously with the adversary.
The researchers have conducted a series of experiments to evaluate the effectiveness of their proposed framework. They tested their approach on nine different text perturbation strategies and five adversarial attack methods, comparing it to several state-of-the-art defense techniques.
The results show that GREATER-A outperforms other methods in terms of semantic preservation, reducing the perturbation rate by up to 10.61% compared to existing defenses. Moreover, GREATER-Adversary is demonstrated to be more effective and efficient than other attack approaches.
To better understand how GREATER-A works, let’s take a closer look at its components. The adversary uses a greedy search algorithm to identify important tokens in the text, which are then perturbed using pruning to create adversarial examples. This process is repeated until a certain threshold is reached or until no further modifications can be made.
Meanwhile, the detector updates its parameters synchronously with the adversary, learning to recognize and defend against the generated adversarial examples. By doing so, GREATER-Detector becomes more robust and able to generalize its defense to different attacks and varying attack intensities.
The researchers have also analyzed the theoretical properties of their proposed framework, providing insights into the perturbation rate and query complexity of the adversarial examples generated by GREATER-A. Their analysis shows that the expected perturbation rate is approximately 7.5%, which is consistent with the experimental results.
Cite this article: “Defending Against Adversarial Attacks in Machine-Generated Text”, The Science Archive, 2025.
Machine-Generated Text, Adversarial Examples, Greater-A, Detector, Adversary, Perturbation, Attacks, Defense, Language Models, Misinformation
