ThesisResearchAI/ML 10-12 months
Adversarial Robustness of Deep Neural Networks
Research on attacks and defences for deep models — generating adversarial examples and evaluating defences like adversarial training and certified robustness.
PythonPyTorchFoolbox / ARTAutoAttack
Find GitHub repos & source code →How to build it — step by step
- 1Attacks: Implement and benchmark attacks (FGSM, PGD, C&W, AutoAttack) on standard datasets.
- 2Defences: Evaluate adversarial training, input transforms, and certified defences.
- 3Robust evaluation: Avoid gradient-masking pitfalls; measure robust accuracy under strong attacks.
- 4Trade-offs: Analyse the robustness-accuracy and compute trade-offs across methods.
Key features to implement
- ✓Multiple attack implementations
- ✓Defence benchmarking
- ✓Rigorous robust evaluation
- ✓Robustness-accuracy trade-off
- ✓Certified-defence study
💡 Unique twist to stand out
Explore transferability of adversarial examples across architectures and propose a defence targeting transfer attacks.
🎓 What you'll learn
Adversarial ML, threat modelling, sound evaluation, and security of ML systems.