sunoh-kim/mac

How sunoh-kim/mac works

This repository implements MAC, a tuning-free test-time defense for CLIP. Its documented workflow evaluates adversarially corrupted images, creates multiple augmented views, applies corruption-aware counterattacks, and aggregates predictions. The graph includes the MAC evaluation code, dataset/model support, and the bundled AutoAttack subsystem. Some relationships rely on the SOURCE INDEX or README; unsampled implementation details are not inferred.

Main entry points: utils.py Evaluation utilities · mac.py MAC evaluation · autoattack.py AutoAttack runner · autopgd_base.py Auto-PGD

Components 26

MAC defense

MAC evaluation code/mac.py
Evaluation utilities code/utils.py

Data and prompts

View augmentation data/datautils.py
Dataset classes code/replace/datasets
Caltech datasets code/replace/datasets/caltech.py
Country211 dataset code/replace/datasets/country211.py
DTD dataset code/replace/datasets/dtd.py
EuroSAT dataset code/replace/datasets/eurosat.py
Aircraft dataset code/replace/datasets/fgvc_aircraft.py
Food101 dataset code/replace/datasets/food101.py
ImageNet variants data/imagenet_variants.py
ImageNet prompts data/imagnet_prompts.py

CLIP models

CLIP loader code/replace/clip.py
CLIP architecture code/replace/model.py
Prompt modules code/models/prompters.py

Attack methods

Attack selection code/attacks.py
Attack helpers code/func.py

AutoAttack

AutoAttack runner autoattack/autoattack.py
Auto-PGD autoattack/autopgd_base.py
FAB attack autoattack/fab_base.py
Square attack autoattack/square.py
Attack checks autoattack/checks.py
Attack utilities autoattack/other_utils.py

Other components

Evaluation user
Prepared datasets
Evaluation results

Connections 31

Evaluation user → MAC evaluation: runs evaluation
MAC evaluation → Attack selection: uses attacks code/mac.py
MAC evaluation → CLIP loader: loads CLIP code/mac.py
MAC evaluation → Prompt modules: uses prompts code/mac.py
MAC evaluation → Evaluation utilities: uses utilities code/mac.py
MAC evaluation → Attack helpers: uses helpers code/mac.py
MAC evaluation → View augmentation: uses augmentation
MAC evaluation → Evaluation results: saves results
Prepared datasets → Evaluation utilities: supplies inputs
Evaluation utilities → View augmentation: builds datasets code/utils.py
Evaluation utilities → ImageNet variants: uses class maps code/utils.py
Evaluation utilities → ImageNet prompts: uses prompts code/utils.py
Evaluation utilities → Caltech datasets: loads dataset code/utils.py
Evaluation utilities → Country211 dataset: loads dataset code/utils.py
Evaluation utilities → DTD dataset: loads dataset code/utils.py
Evaluation utilities → EuroSAT dataset: loads dataset code/utils.py
Evaluation utilities → Aircraft dataset: loads dataset code/utils.py
Evaluation utilities → Food101 dataset: loads dataset code/utils.py
Evaluation utilities → Dataset classes: imports datasets code/utils.py
Dataset classes → Caltech datasets: defines dataset
CLIP loader → CLIP architecture: builds model code/replace/clip.py
Attack selection → AutoAttack runner: invokes runner code/attacks.py
AutoAttack runner → Auto-PGD: dispatches attack autoattack/autoattack.py
AutoAttack runner → FAB attack: dispatches attack autoattack/autoattack.py
AutoAttack runner → Square attack: dispatches attack autoattack/autoattack.py
AutoAttack runner → Attack checks: uses checks autoattack/autoattack.py
AutoAttack runner → Attack utilities: uses utilities autoattack/autoattack.py
Auto-PGD → Attack checks: checks gradients autoattack/autopgd_base.py
Auto-PGD → Attack utilities: uses utilities autoattack/autopgd_base.py
FAB attack → Attack utilities: uses utilities
Square attack → Auto-PGD: references attack autoattack/square.py

Drawn by GitDiagram on November 12, 2026 from the repository's file tree, README and source files. Read as Markdown ·