The rise of Large Language Models (LLMs) enables a new challenge for fake news detection: style-confusion attacks, where attackers rewrite news to mimic trusted styles, challenging detectors that rely on stylistic features. Current defenses like style-invariance training ignore stylistic features, but this is suboptimal for style-confusion attack datasets as it discards valuable cues in clean (non-attacked) articles. To defend against style-confusion attacks, we propose the Robust Adaptive Tri-Encoder Framework (RATF). Our framework uses an LLM to decouple news into content features and style features, and leverages conditional probability curvature to compute a “rewritten score” estimating the likelihood that the article has been rewritten by an LLM. RATF injects this score into a score-conditioned cross-attention fusion module and a learnable parametric gating module. The former uses the content representation as a reliable query to retrieve relevant semantic and stylistic evidence, while the latter continuously calibrates the contributions of original semantics, content features, and style features. We simulate these attacks by creating an equal mixture of original articles from the PolitiFact and GossipCop datasets and their four corresponding style-rewritten versions generated by an LLM. Our experiments show RATF achieves state-of-the-art performance with superior robustness against style-confusion attacks, while also proving highly adaptable across various backbones.

Multi-modal Emotion Recognition in Conversation (MMERC) aims to identify speakers’ emotional states using multi-modal conversational data, significant for various domains. MMERC requires addressing emotional causes: contextual factors that influence emotions, alongside emotional evidence directly expressed in the target utterance. Existing methods primarily model general conversational dependencies, such as sequential utterance relationships or inter-speaker dynamics, but fall short in capturing diverse and detailed emotional causes, including emotional contagion, influences from others, and self-referenced or externally introduced events. To address these limitations, we propose the Evidence-Cause Attention Network for Multi-Modal Emotion Recognition in Conversation (ECERC). ECERC integrates emotional evidence with contextual causes through five stages: Evidence Gating extracts and refines emotional evidence across modalities; Cause Encoding captures causes from conversational context; Evidence-Cause Interaction uses attention to integrate evidence with diverse causes, generating rich candidate features for emotion inference; Feature Gating adaptively weights contributions of candidate features; and Emotion Classification classifies emotions. We evaluate ECERC on two widely used benchmark datasets, IEMOCAP and MELD. Experimental results show that ECERC achieves competitive performance in weighted F1-score and accuracy, demonstrating its effectiveness in MMERC