Authors: Gavad Yash Laxman, Professor Abhishek Popat Bangar, Professor S. B. Bhosale, Professor Karad V. A.
Abstract: Digital documents, paraphrasing tools and large language models have made plagiarism much harder to catch than plain copy-and-paste matching. This review looks at five research papers that approach academic plagiarism detection from different directions: a systematic literature survey, text-based feature selection, deep learning, a comparison of detection techniques, and the detection of AI-obfuscated plagiarism in modeling assignments. Between them, the studies draw on lexical overlap, sentence structure, semantic similarity, writing style and token sequences taken from structured modeling artifacts. One study pairs Chi-square feature selection with a Support Vector Machine (SVM) and reports F-measure scores of 89.34% on PAN 2013 and 92.95% on PAN 2014; another reports that its LSTM-based system placed first among the systems it was compared with on the same benchmarks. A more recent approach for modeling assignments normalizes tokens, matches them automatically and leaves the final judgement to a human inspector, which makes it more resistant to AI-based obfuscation. Taken together, the papers show that no single similarity measure can handle every kind of plagiarism. A practical AI-based checker should therefore combine fast candidate retrieval, several text and structure features, carefully validated machine-learning models and human review. Open problems include resistance to paraphrasing, coverage of datasets and languages, fair evaluation, and the danger of reading a detector score as conclusive proof.
