Neural Deepfake Detection with Factual Structure of Text

Authors: Wanjun Zhong, Duyu Tang, Zenan Xu, Ruize Wang, Nan Duan, Ming Zhou, Jiahai Wang, Jian Yin

Published: 2020-10-15 02:35:31+00:00

Comment: EMNLP2020;10 pages

AI Summary

This paper proposes FAST, a graph-based model for deepfake text detection that leverages the factual structure of documents. It represents the factual structure as an entity graph to learn sentence representations with a graph neural network, then composes them into a document representation for prediction. Experiments show FAST significantly improves detection accuracy over strong transformer-based baselines like RoBERTa by better capturing factual inconsistencies in machine-generated text.

Abstract

Deepfake detection, the task of automatically discriminating machine-generated text, is increasingly critical with recent advances in natural language generative models. Existing approaches to deepfake detection typically represent documents with coarse-grained representations. However, they struggle to capture factual structures of documents, which is a discriminative factor between machine-generated and human-written text according to our statistical analysis. To address this, we propose a graph-based model that utilizes the factual structure of a document for deepfake detection of text. Our approach represents the factual structure of a given document as an entity graph, which is further utilized to learn sentence representations with a graph neural network. Sentence representations are then composed to a document representation for making predictions, where consistent relations between neighboring sentences are sequentially modeled. Results of experiments on two public deepfake datasets show that our approach significantly improves strong base models built with RoBERTa. Model analysis further indicates that our model can distinguish the difference in the factual structure between machine-generated text and human-written text.

Key findings

The FAST model significantly outperforms strong transformer-based baselines, including RoBERTa, BERT, and XLNet, on both news-style and webtext-style deepfake text datasets. The model analysis indicates its ability to distinguish differences in factual structure between machine-generated and human-written text. Ablation studies confirm the individual contributions of factual structure modeling, external knowledge, sequential consistency tracking, and coherence scores to the improved detection accuracy.

Approach

The approach utilizes RoBERTa for contextual word representations, then constructs an entity graph for each document where nodes are entities and edges denote relevance. A multi-layer Graph Convolutional Network (GCN) learns graph-enhanced sentence representations, incorporating external Wikipedia-based entity knowledge. These sentence representations are aggregated into a document representation using an LSTM to track factual consistency and a pre-trained Next Sentence Prediction (NSP) model for coherence, leading to the final deepfake prediction.

Datasets

News-style GROVER-generated dataset, Webtext-style GPT2-generated dataset

Model(s)

UNKNOWN

Author countries

P.R.China

← Previous