<p>Plagiarism is one of the most big issues to deal with in the academic domain due to its effects on the credibility of scientific research. One recognized form of plagiarism involves translating texts from one language into another. Cross-lingual plagiarism is an unethical act that continues to proliferate due to the abundance of online information and the availability of translator tools, especially those based on artificial intelligence. To detect this kind of plagiarism, multilingual transformers offer a promising prospect. In this paper, we propose a new methodology for cross-language plagiarism detection based on multilingual pretrained model mT5 along with a Multi-Head Attention (MHAM) and Attention Mechanisms (AM). The approach includes text preprocessing, embedding using mT5, attention layers, and sigmoid function. To evaluate the efficiency of our approach, we conducted a comparative analysis, demonstrating that our method surpasses other pretrained models in performance, namely XLM-RoBERTa, Multilingual BERT, mBART, and M2M-100. Experiments show that the proposed approach based on mT5 transformer and attention layer achieved the high results for the three language pairs with a plagdet of 98.38% for English-French, 98.03% for English-Spanish, 98.73% for English-German, with a granularity of 1.00 for all language pairs.</p>