Code for our double‑blind submission on securing Concept-based Models (CxMs) under adversarial threat while preserving human intervention effectiveness.
Concept-based Models insert a human-understandable concept layer between a pretrained backbone and the classifier, enabling explanation and human intervention (manual concept correction). Prior work emphasized concept prediction robustness but rarely measured the end-to-end robustness and how interventions interact with adversarial attacks. We evaluate both pre‑intervention and post‑intervention adversarial robustness.
- Install dependencies:
uv sync- Prepare datasets (see Dataset Preparation).
- Create override YAML list (e.g.,
tmp.yaml). - Run an experiment:
python main.py --config tmp.yaml --task_id 0- Omit
--task_idto run all entries.
Defaults reside in class definitions. Override file must be a YAML list; each element merges with defaults (recursive) via utils.yaml_merge. The merged config and checkpoint are stored under checkpoints/; run logs go to logs/<run_name>/<run_id>/ with deterministic run_name + hashed run_id.
Example cfg.yaml:
- model: CBM
dataset: CUB
train_mode: Std
# below is config of RAIDCXM
- model: VCBM
dataset: CUB
train_mode: JPGD
res_dim: 32
hsic_weight: 0.0001Run the second entry:
python main.py --config cfg.yaml --task_id 1Expected structure under data/:
CUB_200_2011/
images/ ...
attributes/
train.pkl / val.pkl / test.pkl
Animals_with_Attributes2/
JPEGImages/ ...
classes.txt
predicates.txt
predicate-matrix-binary.txt
Ensure data_path points to the parent directory. To add a dataset, implement a loader class under dataset/.
For CUB, *.pkl files can be found in CBM's repository here.