HAT4RD: Hierarchical Adversarial Training for Rumor Detection in Social Media

Area of Science:

• Computer Science
• Social Media Analysis
• Artificial Intelligence

Background:

• Social media facilitates communication but also enables rapid rumor dissemination.
• Rumors can negatively impact public judgment and social security.
• Existing rumor detection models struggle with the high-dimensional and sparse nature of natural language, limiting their robustness and generalization.

Purpose of the Study:

• To propose a novel hierarchical adversarial training method for rumor detection (HAT4RD) on social media.
• To enhance the robustness and generalization capabilities of rumor detection models.
• To address the challenge of diverse rumor expressions on social media.

Main Methods:

• Developed a hierarchical adversarial training (HAT4RD) approach incorporating gradient ascent with adversarial perturbations.
• Applied perturbations to embedding layers at both post-level and event-level modules.
• Employed stochastic gradient descent for the detector to minimize adversarial risk and learn robust representations.

Main Results:

• Verified the model's robustness against various adversarial attacks.
• Demonstrated improved generalization through visual experiments showing a drift towards a flat loss landscape.
• Achieved superior performance compared to state-of-the-art methods on public datasets from Twitter and Weibo.

Conclusions:

• HAT4RD significantly enhances the robustness and generalization of social media rumor detection models.
• The adversarial training strategy effectively addresses the challenge of diverse rumor expressions.
• The proposed method represents a significant advancement in combating online misinformation.