adversarial-ai-target
EfficientNet-B3 fine-tuned for binary chest X-ray classification. Built as the primary attack target for the adversarial-ai-attacks-mitigations research series.
Model Details
| Property | Value |
|---|---|
| Architecture | EfficientNet-B3 (ImageNet pretrained) |
| Task | Binary image classification |
| Classes | NORMAL (0), PNEUMONIA (1) |
| Input size | 300 × 300 RGB |
| Framework | PyTorch 2.0 |
| Dataset | Kaggle chest-xray-pneumonia |
Training
| Property | Value |
|---|---|
| Phase 1 (epochs 1-4) | Backbone frozen, head only, lr=1e-3 |
| Phase 2 (epochs 5-10) | Last 3 backbone blocks unfrozen, lr=1e-4 |
| Optimizer | AdamW |
| Scheduler | CosineAnnealingLR |
| Batch size | 64 (A100) |
| Class balancing | WeightedRandomSampler |
Performance
| Metric | Value |
|---|---|
| Test Accuracy | 0.8862 |
| AUC | 0.9738 |
| PNEUMONIA Recall | 0.99 |
| NORMAL Precision | 0.99 |
Intended Use
This model is intended strictly for adversarial AI security research and education. It serves as the attack surface for chapters 4-9 and 12 of the hands-on lab series covering poisoning attacks, evasion attacks, model extraction, membership inference, and GAN-based attacks.
Do not use this model for clinical decision making.
Research Series
Part of The Inference Loop research series.