Case study
DiffusionOOD
Conditional diffusion models used as generative classifiers for OOD detection, with separation loss reaching 99.03% +/- 0.07% AUROC.
Problem
Build a more stable OOD detector that does not rely on classifier confidence alone.
Role
Master thesis author, modeling, experiments, evaluation, and analysis.
Approach
- Train a binary conditional diffusion model and score reconstruction error under competing class conditions.
- Add class-conditional separation loss to push conditional noise predictions apart.
- Evaluate with multiple seeds and external zero-shot benchmarks instead of relying on one lucky run.
Results
- 99.03% +/- 0.07% AUROC averaged over three seeds on CIFAR-10.
- +6.5 percentage points over the non-separated baseline.
- Seed-42 generalized zero-shot to CIFAR-100, Places365, FashionMNIST, Textures, and SVHN.
Lessons
- Stability matters almost as much as peak score.
- A result becomes more useful when variance is reported honestly.
- Cross-domain transfer should be tested, not assumed.