AI-SPAMGUARD: A REPRODUCIBLE MULTI-TRANSFORMER AND EXPLAINABLE DECISION-FUSION FRAMEWORK FOR ROBUST DETECTION OF LLM-GENERATED OPINION SPAM

Abstract

Large language models (LLMs) can generate persuasive online reviews at negligible marginal cost, raising a practical threat to review-platform integrity. This paper presents AI-SpamGuard, a reproducible multi-transformer evaluation and decision-fusion framework for detecting LLM-generated hotel opinion spam. The study reconstructs and extends the LLM-review benchmark of Liyanage et al. using validated human reviews together with GPT-3, LLaMA, and paraphrased AI-generated reviews. After data auditing, the usable corpus contains 800 authentic reviews, 800 GPT-3 reviews, 790 LLaMA reviews, and 800 paraphrased reviews. Four transformer detectors—BERT, SciBERT, ELECTRA, and XLNet—are fine-tuned for 10 epochs with three independent seeds. Native-tokenizer mean generated-class F1 reaches 99.02% on GPT-3 (BERT), 96.76% on LLaMA (XLNet), and 99.28% on paraphrased reviews (ELECTRA). A post-hoc majority-vote decision layer computed from synchronized saved predictions improves LLaMA F1 to 97.04% and paraphrased F1 to 99.30%, although it is slightly below BERT on GPT-3, demonstrating that fusion is generator-dependent rather than universally beneficial. Robustness tests reveal the central limitation: strong in-domain performance does not guarantee cross-generator transfer. For example, the strong extension detector obtains only 52.21% generated-F1 when trained on GPT-3 and tested on LLaMA, while unseen-hotel evaluation remains substantially stronger. Explainability is provided through the reproducible TF-IDF/Multinomial-Naive-Bayes baseline using additive lexical attribution and a locality-weighted LIME-style surrogate. The paper also reports aggregate confusion matrices, probability ROC-AUC for the saved baseline models, tokenizer ablation, leave-one-generator-out evaluation, and reproducibility constraints. The results support a practical conclusion: detector validation must emphasize generator shift and protocol fidelity, not only near-saturated in-domain accuracy.