AB
AB
AiBoss

All Categories

project

ImBD - A general-purpose AI content detector jointly launched by Fudan University, South China University of Technology, and other institutions.

ImBD (Imitate Before Detect) is a method developed by Fudan University, South China University of Technology, Wuhan University, Fenzi AI, and others for detecting machine-edited text. It first mimics the text generation process of large language models (LLMs)...

What is ImBD?

ImBD (Imitate Before Detect) is a method developed by Fudan University, South China University of Technology, Wuhan University, Fenzi AI, and others for detecting machine-edited text. First, it mimics the style of text generated by large language models (LLMs), adjusting the scoring model based on style preference optimization (SPO) to better match the characteristics of machine-edited text. Then, it utilizes style conditional probability curvature (Style-CPC) to quantify the log probability difference between the original text and the conditionally sampled generated text, effectively distinguishing between human writing and machine-edited content. ImBD performs exceptionally well in various scenarios, including text revisions by different LLMs, multiple text domains, and various revision types, significantly improving detection performance, and achieving superior performance with only a small amount of data and a short training time.

Main functions of ImBD

  • Detection machine revised textEffectively identifies machine-edited text, including rewrites, expansions, and polishing. Captures unique machine-style features in text, distinguishing between human writing and machine-edited content, thus improving detection accuracy.
  • Adaptable to various scenariosImBD is applicable to various text domains and different types of machine editing tasks. Including news, academic paper writing, and other fields, ImBD can accurately detect traces of machine editing and has good generalization capabilities.
  • Efficient Training and ReasoningImBD achieves superior performance with only a small number of samples and a short training time. It rapidly processes text data during the inference phase, providing detection results and meeting the efficiency requirements of practical applications.

ImBD's technical principles

  • Style Preference Optimization (SPO)The scoring model was adjusted based on SPO to better align with the style characteristics of machine-revised text. Using pairs of human-written and machine-revised texts, the token distribution of the optimized model was made more aligned with the style of the machine-revised text, enhancing the model's ability to recognize machine style features.
  • Style Conditional Probability Curvature (Style-CPC)During the detection phase, I uses Style-CPC to quantify the log probability difference between the original text and the conditionally sampled generated text. Based on this difference, the model can effectively distinguish between human-written and machine-edited content, because machine-edited text typically exhibits different characteristics in its probability distribution compared to human-written text.
  • Combining Imitation and DetectionThe core idea of ImBD is to first mimic the machine's style before performing detection. By mimicking the machine's style characteristics, the model can better understand the features of machine-revised text, more accurately identify traces of machine revisions during detection, and improve the accuracy and robustness of detection.

ImBD project address

Application scenarios of ImBD

  • academic fieldIt helps reviewers identify machine-edited parts of papers, ensuring originality and academic integrity.
  • News mediaDetects machine-edited content in news articles to ensure the authenticity and accuracy of reporting.
  • Publishing industryIt assists editors in proofreading manuscripts, identifies machine-edited revisions, and ensures book quality.
  • EducationIt assesses the machine-assisted portion of student assignments, accurately evaluates student abilities, and provides targeted feedback.
  • Enterprise and BusinessReview marketing copy and other materials to ensure originality and quality, and avoid copyright and brand issues.
  • Network security and information verificationTo identify machine-generated or modified content in online misinformation and maintain network security and information authenticity.
Back