[CVPR 2026] The Power of Decaying Steps: Enhancing Attack Stability and Transferability for Sign-based Optimizers.

---

🚀 Environment Setup

Since our mechanisms are built upon established frameworks, please follow the original setup instructions:

• Image Classification: Refer to the official TransferAttack repository for dependency and dataset setup.
• Cross-modal Retrieval: Refer to the official SA-AET repository for multi-modal environment setup (e.g., MS-COCO configuration, Transformers).

---

🛠️ Usage

TransferAttack (Image Classification)

Running an Attack Example: Assuming you are in the TransferAttack root directory with properly configured datasets, run:

# Generate adversarial examples using MDCS-OPS
python main.py --input_dir ./data --output_dir adv_data/mdcsmi/resnet50 --attack mdcsmi --model resnet50

# Evaluate the generated adversarial examples
python main.py --input_dir ./data --output_dir adv_data/mdcsmi/resnet50 --eval

Retrieval (Cross-modal Retrieval)

The Retrieval folder contains the plus implementations which denote our integration of the MDCS step-size adaptive mechanism into existing baselines:

• SA_AET_plus.py
• SGAttacker_plus.py
• DRAttacker_plus.py

Running an Attack Example: Assuming you are in the Retrieval root directory with properly configured datasets, run:

# Evaluate the generated adversarial examples on SGAttacker + MDCS
python eval.py --config ./configs/Retrieval_flickr.yaml --attack mdcssga

---

🙏 Acknowledgements

We would like to thank the authors of the following outstanding open-source repositories, which significantly facilitated our research and codebase construction:

• TransferAttack
• SA-AET
• VLPTransferAttack