<p>The Google Play Store amasses user opinions through star ratings and written reviews, offering rich insights into satisfaction that guide download decisions. Sentiment analysis, an NLP technique, enables machine-learning models to quantify the emotions contained in these reviews. Yet recent work shows that deep-learning models supporting such analysis are susceptible to adversarial attacks—subtle textual tweaks invisible to humans but capable of driving misclassification. This study introduces Text-Muddler, a novel black-box, word-level adversarial attack that inserts minimal, semantically coherent perturbations to fool target models while preserving human readability. Unlike prevailing word-level methods such as TextFooler and PWWS, Text-Muddler operates beneath the tokenizer threshold, rendering its distortions extremely hard to filter. We evaluated Text-Muddler against four widely used transformer architectures—BERT, DistilBERT, ALBERT, and RoBERTa—each fine-tuned on a corpus of 25,000 reviews drawn from five productivity applications on the Google Play Store. Attack effectiveness was quantified through attack success rate (ASR), reflecting the proportion of samples whose predicted sentiment flipped after perturbation. Experimental results show that Text-Muddler surpasses state-of-the-art attacks, achieving higher ASR with fewer character modifications. A comparative robustness analysis uncovers marked disparities among transformers, revealing that some architectures are substantially more resilient than others to Text-Muddler’s manipulations. These findings underscore the fragility of sentiment-analysis pipelines in real-world marketplaces and highlight an urgent need for defensive strategies—such as adversarial training, input sanitization, and continual model monitoring—as well as regulatory oversight to curb malicious exploitation and misinformation spread. Future work will investigate adaptive defenses and benchmark Text-Muddler across additional languages and domains in depth.</p>