Breaking the Adversarial Robustness-Performance Trade-off in Text Classification via Manifold Purification

The introduction of the Manifold-Correcting Causal Flow (MC^2F) represents a significant advancement in the field of text classification, particularly in overcoming the longstanding adversarial robustness-performance trade-off. Traditional approaches often compromise clean data performance to enhance robustness against adversarial attacks. However, MC^2F utilizes a two-module system that models the distribution of clean samples in the encoder embedding manifold, effectively correcting out-of-distribution embeddings. Extensive evaluations across three datasets demonstrated that this method not only establishes a new state-of-the-art in adversarial robustness but also fully preserves performance on clean data, with modest gains in accuracy. This breakthrough is crucial for developing more reliable text classification systems that can withstand adversarial challenges while maintaining high performance, thus enhancing their applicability in various real-world scenarios.