Research code accompanying a master's thesis at the University of Belgrade, School of Electrical Engineering. The study investigates NFT scam detection through three independently evaluated sources of evidence: images, textual descriptions, and collection website domains.
Author: Stefan Veličković · Advisor: Prof. Dr. Boris Delibašić · Date: September 2023
Publication status: Thesis in the process of being published.
Original title: Примена вештачке интелигенције за препознавање лажних и малициозних NFT-ева
Research overview · Notebooks · Results · Reproducibility · Citation
NFT scams can exploit visual presentation, descriptive text, and collection websites. This thesis examines how different machine learning approaches use each of those signals to distinguish examples labeled legitimate from examples labeled scam or malicious.
The work compares classical classifiers, convolutional and recurrent neural networks, transfer learning, and a transformer-based text classifier. The three notebooks are separate experiments; the repository does not implement a trained end-to-end fusion model combining all modalities.
The research provides:
- An image-classification study comparing a custom CNN, ResNet50, VGG16, an LSTM-based network, and a ResNet50/CNN hybrid.
- A description-classification study covering TF-IDF, Word2Vec, classical machine learning, DistilBERT, and a feedforward network.
- A collection-domain study using lexical features, TF-IDF, tree-based methods, class-imbalance techniques, and recurrent models.
- Preserved notebook outputs documenting the original exploratory work, alongside a guide to the reported results and reproduction constraints.
| Study | Notebook | Main approaches |
|---|---|---|
| NFT images | Images_NFTs_Master.ipynb | CNN, ResNet50, VGG16, bidirectional LSTM, hybrid network |
| NFT descriptions | Metadata_NFTs_Master.ipynb | TF-IDF, Word2Vec, classical classifiers, DistilBERT, FNN |
| Collection domains | Domains_NFT_Master.ipynb | Lexical features, TF-IDF, Random Forest, SMOTE, LightGBM, XGBoost, recurrent networks |
The notebooks include saved experiment outputs and links to open each study in Colab.
The thesis reports the following highest accuracies within each study:
| Input modality | Highest-accuracy model reported in the thesis | Accuracy | Thesis reference |
|---|---|---|---|
| Images | Custom CNN | 98.36% | Table 4, printed page 57 |
| Descriptions | Random Forest | 94.83% | Table 5, printed page 59 |
| Collection domains | XGBoost | 84.32% | Table 6, printed page 61 |
These figures are reported thesis results, not new measurements or independently reproduced benchmarks. Some saved notebook outputs differ from the thesis tables, and the modalities use different datasets and evaluation setups. Results and provenance records those distinctions; the percentages should not be treated as a controlled comparison of modalities.
The thesis describes data collection from Phantom's blocklists, NFT marketplaces, and a Kaggle collection dataset. The research datasets are not distributed in this repository. To discuss data access and reproducing the experiments, contact Stefan Veličković.
Labels are derived from the collection and blocklist sources used for the study. They are research labels, rather than independent guarantees about a token, creator, or collection. NFT images and third-party metadata remain subject to their respective owners' rights.
See Data documentation for expected inputs, manuscript counts, notebook counts, and known gaps.
Open the notebooks directly on GitHub to inspect their saved outputs. For execution, arrange data access with the author and set NFT_DATA_ROOT using the documented layout.
A complete historical dependency lock was not captured. The notebooks use legacy Keras preprocessing APIs and several scientific Python frameworks; installing the latest versions of every package is not a verified reproduction procedure. The reproduction guide explains the limits and requirements for future execution.
Lightweight repository checks can be run without downloading datasets or training models:
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements-test.txt
python scripts/validate_repository.py
python -m unittest discover -s tests -v
cffconvert --validateOn Windows, activate the environment with .venv\Scripts\activate. These checks validate notebook syntax, historical-output preservation, documentation, citations, and selected code paths on synthetic inputs. They do not train models or reproduce research scores. Optional framework tests skip when TensorFlow/Pillow is absent.
The study is limited by the number and provenance of scam examples, class imbalance, incomplete metadata, and the changing nature of NFT scams. Stronger follow-up evaluation should include collection-disjoint and time-based splits, duplicate control, training-only feature fitting, probability-based ROC-AUC, repeated trials, and uncertainty estimates. These are directions for further validation, not experiments claimed to have been completed here.
Research notes separates limitations documented in the thesis from observations about the archived notebooks.
This thesis research provided the basis for MintEye, a later Solana Radar Hackathon application for NFT risk classification. The thesis repository preserves the research experiments; MintEye contains the subsequent application and inference workflow. Its models and runtime should not be treated as the exact thesis evaluation artifacts.
For the thesis:
Veličković, S. (2023). Примена вештачке интелигенције за препознавање лажних и малициозних NFT-ева [Unpublished master's thesis, University of Belgrade, School of Electrical Engineering].
The English heading above is a descriptive translation. No DOI or publication venue is asserted. Use CITATION.cff for machine-readable citation metadata, or citation.bib for BibTeX. The full thesis manuscript is not redistributed in this repository.
Images_NFTs_Master.ipynb Image experiments and saved outputs
Metadata_NFTs_Master.ipynb Description experiments and saved outputs
Domains_NFT_Master.ipynb Domain experiments and saved outputs
docs/ Data, results, methods, and reproduction notes
assets/ Repository presentation assets
scripts/ Notebook syntax and repository validation
tests/ Offline synthetic regression checks
CITATION.cff Machine-readable research citation
citation.bib BibTeX thesis citation
LICENSE MIT license for repository code
The repository code is available under the MIT License. This does not grant rights to third-party NFT artwork, source datasets, or the separately supplied thesis manuscript.
