CSEG3060 · Research Methodology in Computer Science
Definition, types — direct, self, mosaic, accidental — and detection tools
Attribution is not a formatting requirement. It is the mechanism by which knowledge remains traceable.
Unit III · Lecture 4 of 8 · 60 minutes · CO3
Dr. Mohsin Furkh Dar · School of Computer Science, UPES
→ / Space next · ← back · O overview · N speaker notes · F full screen
Learning objectives · CO3
Define plagiarism in academic and research contexts with precision
Differentiate the major types: direct, self, mosaic and accidental
Apply citation frameworks — APA, IEEE, ACM, MLA — to mitigate attribution errors
Interpret similarity reports from detection tools critically
Recognise the ethical, legal and professional consequences in CS research
Formulate institutional and personal strategies to uphold integrity
Context
A 2023 Retraction Watch analysis indicates plagiarism-related retractions in computer science rose approximately 38% over the preceding decade.
A significant proportion is attributed to inadequate paraphrasing and improper reuse of prior work — not to brazen copying.
Source code
Algorithms
Datasets
Documentation
The two fastest-growing categories — poor paraphrase and self-reuse — are precisely the ones students believe are not plagiarism.
Definition
From Latin plagiarius — “kidnapper”.
| Source | Definition |
|---|---|
| ORI (USA) | “the appropriation of another person's ideas, processes, results, or words without giving appropriate credit” |
| Council of Science Editors | using the work, ideas or words of others without proper attribution, irrespective of intent |
| IEEE Publication Policy | “copying ideas, text, data, software, or any other material without proper citation” |
| UGC (India) Regulations 2018 | taking someone else's work or ideas and passing them off as one's own |
The material originates from a verifiable source
Credit not commensurate with what was borrowed
Deliberate deception or reckless disregard
From a phrase to an entire manuscript
Type 1 of 4
“Deep learning is a subset of machine learning that employs neural networks with many layers. These networks can learn complex patterns in large datasets, enabling breakthroughs in image recognition, natural language processing, and autonomous systems.”
Textbook-level content a student may consider common knowledge — yet the specific wording belongs to its author. Quotation marks plus a citation converts the offence into scholarship.
A directly copied paragraph typically exceeds 85–90% overlap within that segment — which is why a document-level index of 12% can still conceal a wholly copied paragraph.
Type 2 of 4
A 2022 paper in Journal A on “Efficient Graph Neural Networks for Social Networks”, substantially reproduced in a 2024 submission to Conference B without citing the earlier work — self-plagiarism, notwithstanding common authorship.
The legitimate case: a conference paper extended into a journal article is entirely acceptable when the extension is disclosed, the earlier paper cited, and the added contribution stated.
Type 3 of 4
“Convolutional neural networks have revolutionized image classification by automatically learning hierarchical features from raw pixel data.”
“CNNs have transformed image categorization by autonomously extracting multi-level representations directly from unprocessed pixel inputs.”
Vectors u and v represent the original and submitted passages. Values above roughly 0.80 signal significant semantic overlap. Near-parallel vectors mean near-identical meaning regardless of surface wording.
Type 4 of 4
Reproducing information without having recorded the reference.
Insufficient transformation, with no quotation marks.
Following a source's organisational outline too closely.
Subconscious reproduction of material believed to be original.
Citing a secondary source in place of the original, or the reverse.
A student summarises a Wikipedia paragraph, omits the citation because they “thought it was common knowledge”, and submits it. Accidental — and nonetheless plagiarism under most institutional policies.
Because the causes are procedural, the fixes are procedural: a literature log, a reference manager, and citation capture at the moment of reading rather than at the moment of writing.
Emerging and CS-specific
| Form | Description |
|---|---|
| Idea plagiarism | Appropriating concepts or theoretical frameworks without attribution |
| Code plagiarism | Copying source code, algorithms or software components without credit — acutely relevant in CS |
| Data plagiarism | Reusing datasets without permission or acknowledgement |
| Image plagiarism | Reproducing figures, diagrams or screenshots without citation |
| Translation plagiarism | Translating a foreign-language source and presenting it as original |
| AI-assisted plagiarism | Submitting AI-generated text as one's own, or paraphrasing AI output without disclosure |
Open-source licences — MIT, GPL, Apache, BSD — impose specific attribution formats. Copying a Stack Overflow snippet or a GitHub function without the required notice is simultaneously an academic and a licensing violation.
Venues differ, but the emerging norm is disclosure of AI assistance plus author responsibility for verification. Unverified AI output additionally risks fabricated citations.
Detection
| Approach | Methodology | Strengths | Limitations |
|---|---|---|---|
| String matching | Exact or near-exact text comparison | Fast, transparent | Fails on paraphrased content |
| Fingerprinting | Hash-based indexing of n-grams | Efficient over large corpora | Sensitive to word-order change |
| Stylometry | Statistical analysis of writing style | Detects ghostwriting | Needs substantial reference text |
| Semantic similarity | NLP embedding comparison | Identifies paraphrase | Computationally intensive |
| Citation analysis | Examination of citation patterns | Detects missing citations | Limited to structured documents |
String matching catches direct plagiarism · fingerprinting scales it to web-sized corpora · semantic similarity catches mosaic plagiarism · stylometry addresses authorship, not overlap.
Identifier renaming, statement reordering and reformatting defeat text comparison entirely — which is why token-based and structural methods exist for source code.
Prevention
Practical application
Turnitin — institutional student submissions; web, student-paper repositories and publications; similarity index with highlighted matches; LMS integration. Weak on well-paraphrased text and ideas.
iThenticate — publisher-side, built on the CrossCheck database, integrated with editorial workflow. Subscription; primarily textual.
Grammarly · Copyscape · PlagScan · Urkund/Ouriginal — lightweight, web-oriented or institutional; smaller databases.
MOSS — Measure of Software Similarity: purpose-built for code, using winnowing and fingerprinting adapted to programming structure. The standard for programming assignments.
JPlag — open-source, multi-language, token-based comparison.
GPTZero · Originality.ai · Turnitin AI Writing Detection · Copyleaks — estimate machine generation from perplexity, burstiness and stylistic pattern.
Critical caveat. AI-detector reliability remains contested, with documented false-positive rates — non-native English writers and heavily edited prose are disproportionately flagged. Such output is a prompt for human inquiry, never evidence of misconduct on its own.
Practical application
Thresholds are institution- and journal-specific. IEEE permits up to 30% for review articles under certain conditions; the UGC Regulations 2018 set graded levels for theses with escalating consequences, beginning at a tolerance of up to 10%.
32% can be clean — if composed of quoted, cited and bibliographic matches.
6% can be serious — if that 6% is an uncited verbatim paragraph.
Case studies
Hwang Woo-suk (2004–05). Landmark Science papers claiming stem-cell cloning; investigation revealed fabricated data and duplicated figures, with image reuse and uncredited methodology. Retraction, revoked funding, criminal conviction. Lesson: applying detection to figures and data reveals what textual review misses.
Springer retraction (2019). A CS paper retracted after editors identified substantially overlapping content in two venues without cross-citation. Lesson: self-plagiarism is serious even where the author owns both works.
Code plagiarism in programming courses. MOSS-based studies report 15–25% of submitted assignments containing significant overlap in some cohorts. Mitigations: MOSS or JPlag screening, per-student parameterisation, viva voce verification, integrity education from week one.
Zero grade · course failure · probation · suspension · expulsion · degree revocation
Retraction · journal and conference blacklisting · loss of funding · termination
Copyright infringement suits · breach of publisher agreements · damages and injunctions
Durable loss of credibility · difficulty securing collaborations · a retraction notice is a permanent, indexed, citable record
Consolidation
Work this one. A manuscript reuses four paragraphs verbatim from the authors' own 2021 conference paper with neither citation nor quotation marks, and paraphrases two further paragraphs from a 2019 journal article without attribution. Classify each instance and prescribe the corrective action.
Next — Lecture 5: Plagiarism Prevention Strategies: paraphrasing, citation styles, reference management and Turnitin/URKUND usage.