Models

MedDef model variants and their training stages (6 variants × 2 stages)

Primary

Primary deployment model · full DAAM architecture (DISTILL_V2 final)

Stage	Description	Status	Accuracy	ONNX Path
Stage 1	Full-parameter supervised training	Pending export	—	not exported
Distill v2 (Final)	Final model · enhanced distillation with adversarial fine-tuning	Pending export	—	not exported

Complete model with adversarial defense module enabled

Stage	Description	Status	Accuracy	ONNX Path
Stage 1	Full-parameter supervised training	Pending export	—	not exported
Distill v2 (Final)	Final model · enhanced distillation with adversarial fine-tuning	Pending export	—	not exported

Ablation: no dual-frequency attention branch

Stage	Description	Status	Accuracy	ONNX Path
Stage 1	Full-parameter supervised training	Pending export	—	not exported
Distill v2 (Final)	Final model · enhanced distillation with adversarial fine-tuning	Pending export	—	not exported

Ablation: no patch attention module

Stage	Description	Status	Accuracy	ONNX Path
Stage 1	Full-parameter supervised training	Pending export	—	not exported
Distill v2 (Final)	Final model · enhanced distillation with adversarial fine-tuning	Pending export	—	not exported

Ablation: no CBAM channel attention

Stage	Description	Status	Accuracy	ONNX Path
Stage 1	Full-parameter supervised training	Pending export	—	not exported
Distill v2 (Final)	Final model · enhanced distillation with adversarial fine-tuning	Pending export	—	not exported

Standard ResNet-18 without any attention or defense

Stage	Description	Status	Accuracy	ONNX Path
Stage 1	Full-parameter supervised training	Pending export	—	not exported
Distill v2 (Final)	Final model · enhanced distillation with adversarial fine-tuning	Pending export	—	not exported