A reproducible PyTorch lab for evaluating FGSM evasion attacks and adversarial training on Fashion-MNIST.
- Standard and adversarially trained CNNs
- White-box FGSM evaluation with epsilon 0.15
- Reproducible CSV, chart, image, and log evidence
- Streamlit interface for interactive comparison
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe -m src.train_standard
.\.venv\Scripts\python.exe -m src.train_adversarial
.\.venv\Scripts\python.exe -m src.evaluate
.\.venv\Scripts\python.exe -m streamlit run app\streamlit_app.pyReview the repository files before running commands that access external services or download models and datasets. Never commit credentials or personal data.
This is an educational or experimental project. See ROADMAP.md for intentionally scoped future work and CONTRIBUTING.md before proposing changes.
Original source code is available under the MIT License. Third-party datasets, models, media, services, course materials, and dependencies keep their own terms; see THIRD_PARTY_NOTICE.md.
Laboratório reproduzível em PyTorch para comparar uma CNN padrão e uma CNN com treinamento adversarial diante de ataques FGSM white-box no Fashion-MNIST. Exporta evidências e inclui demonstração em Streamlit.
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe -m src.train_standard
.\.venv\Scripts\python.exe -m src.train_adversarial
.\.venv\Scripts\python.exe -m src.evaluate
.\.venv\Scripts\python.exe -m streamlit run app\streamlit_app.pyConsulte o ROADMAP.md para funcionalidades futuras e o CONTRIBUTING.md antes de contribuir. Nunca envie credenciais ou dados pessoais ao repositório.
O código-fonte original é disponibilizado sob a Licença MIT. Datasets, modelos, mídias, serviços, materiais acadêmicos e dependências de terceiros preservam seus próprios termos; consulte THIRD_PARTY_NOTICE.md.