Is Multi-Modal Necessarily Better? Robustness Evaluation of Multi-Modal Fake News Detection

The proliferation of fake news and its serious negative social influence push detection methods to become necessary tools for web managers. Meanwhile, the multi-media nature media makes multi-modal popular ability capture more modal features than uni-modal methods. However, current literature on is likely pursue accuracy but ignore robustness (the in case abnormality malicious attack) detector. To address this problem, we propose a comprehensive evaluation detectors. In work, simulate attack users developers, i.e., posting injecting backdoors. Specifically, evaluate detectors with five adversarial two backdoor Experiment results imply that: (1) performance state-of-the-art degrades significantly under attacks, e.g., BDANN's drops by 47% compared normal, even worse general (Att-RNN); (2) Most multimodal are vulnerable visual modality textual modality; (3) Backdoor attacks events severely degrade (accuracy dropped an average 20%); (4) These (another 2% reduction accuracy) when subjected attacks; (5) Defense will improve detectors, cannot fully resist effects attacks.