Abstract
<title>Abstract</title> <p>Fake information detection remains vulnerable to label-preserving surface variation: a false claim can retain its veracity status while being reformulated through lexical noise, neutral paraphrasing, credibility-cue insertion, or stylistic camouflage. Clean-set performance therefore provides an incomplete estimate of reliability, because it does not reveal whether a detector relies on stable veracity evidence or brittle surface correlations. This paper proposes SPAT-UAD, a robust text-based detection framework that couples semantics-preserving adversarial training with uncertainty-aware adaptive decision. During training, each claim is paired with multiple label-preserving views and optimized through clean classification, adversarial-view classification, clean–perturbed consistency, and calibration objectives. During inference, calibrated confidence, entropy, probability margin, and view agreement are converted into an instance-specific decision threshold. Experiments on LIAR, Constraint-COVID, and PHEME5 demonstrate that SPAT-UAD achieves the strongest robust score (74.4), average heavy-condition Macro-F1 (68.2), AUROC (88.6), and expected calibration error (0.054) among fourteen text-compatible baselines. Compared with the strongest baseline, the proposed method improves robust score by 5.4 points and heavy-condition Macro-F1 by 7.5 points. Ablation, calibration, degradation, and qualitative analyses indicate that the improvement is driven by perturbation-invariant representation learning, semantic-distractor resistance, multi-channel textual evidence, probability calibration, and uncertainty-aware threshold control.</p>