Adversarial learning for fake news detection

Adversarial learning encompasses multiple approaches where systems are trained or evaluated in hostile/competitive settings to improve robustness.

Domain-invariant feature learning: Uses a minimax game between two neural networks to improve feature learning. In the context of fake news detection, adversarial approaches address learning features that generalize across different domains or conditions (e.g., different events, different news sources, different time periods). A feature extractor is trained to simultaneously: 1. Minimize a primary task loss (e.g., fake news classification) 2. Maximize an adversarial discriminator loss (e.g., the discriminator cannot identify which event, source, or domain the features come from)

This forces the feature extractor to learn representations that are invariant to domain-specific variations while remaining discriminative for the target task.

Adversarial robustness testing: Generating adversarial samples (evasion examples or simulated future attacks) to test detection system vulnerabilities and inform defensive improvements. Rather than training with an adversary, this approach proactively simulates likely attack evolutions (e.g., bot evolution) to discover and address detector weaknesses before attackers do.

Key papers

• Better Safe Than Sorry: an Adversarial Approach to improve Social Bot Detection — Proposes GenBot, a genetic algorithm for adversarially simulating evolved spambots; generates adversarial samples (evolved bots) that evade state-of-the-art detection systems, revealing vulnerabilities and suggesting improvements (entropy-based features); exemplifies proactive adversarial robustness approach: anticipate future attacks to preemptively strengthen defenses
• Wang et al. (2018) — EANN: Proposes Event Adversarial Neural Networks where an event discriminator forces the multi-modal feature extractor to learn event-invariant representations. Uses gradient reversal layer to implement the minimax game. Demonstrates 10.3% improvement on Twitter dataset over prior multi-modal methods by removing event-specific features that don't transfer to new events.

Connections

• Transfer learning is the broader problem: adversarial learning is one technique for achieving domain-invariant features.
• Deep learning provides the neural network architectures underlying adversarial training.
• Related to domain adaptation literature, where gradient reversal and adversarial discriminators are widely used.