Competition on Evolutionary Computation in Multi-Label Adversarial Examples

This repository provides the benchmark framework and source code for the Competition on Evolutionary Computation in Multi-Label Adversarial Examples.

Competition Results and Submitted Algorithms

The final competition results and the source code of the three submitted algorithms are available in organized_submissions/. That directory contains the final ranking, the eight benchmark results for each entrant, English algorithm descriptions, and the organized implementation of each submission.

Final ranking:

1. Juewen Fan - Gradient-Plateau SBA
2. Chaofan Xie - Multi-Archive Search with SBA
3. Gaoren Zhang - Structured Active-Label Attack

---

📁 Dataset Preparation

Please download the MLAE_cec_data folder from the following link:

🔗 Download: https://drive.google.com/file/d/1mTPTDYoMzpcCOxeuiZfOPkxpFYdk9zcV/view?usp=drive_link

After downloading, place the folder anywhere on your machine and specify its full path in proConfig["data_dir"].

Directory Structure

MLAE_cec_data
│
├── adv_filter_data
│   ├── coco
│   ├── nuswide
│   ├── voc2007
│   └── voc2012
│
└── checkpoint
    ├── mlgcn
    └── mlliw

---

⚙️ Usage Example

Below is a minimal example for running an evolutionary adversarial attack:

import random
from attack_problem import EvolutionaryAttackProblem
from de import DE_RAND1

proConfig = {
    "ml_model_name": "mlgcn",        # mlgcn / mlliw
    "dataset_name": "coco",          # coco / voc2007 / voc2012 / nuswide
    "target_type": "random",
    "epsilon": 77.596,               # Perturbation limit ε
    "max_eval": 10000,               # Max fitness evaluations per image (query limit)
    "data_dir": "/home/dyy/code/MLAE_cec_data"  # Path to MLAE_cec_data
}

algConfig = {
    "F": 0.5,
    "CR": 0.9,
    "pop_size": 100,
    "rnd": random.Random(1234)
}

problem = EvolutionaryAttackProblem(proConfig)
mlde = DE_RAND1(algConfig)

problem.attack(mlde)
print("attack_rate:", problem.attack_rate())

---

🔧 Configuration Details

Problem Configuration (proConfig)

Key	Description
ml_model_name	Multi-label model to attack: mlgcn, mlliw
dataset_name	Dataset: coco, voc2007, voc2012, nuswide
target_type	Attack types (random)
epsilon	L2 perturbation bound
max_eval	Maximum number of fitness evaluations (query limit)
data_dir	Local path to MLAE_cec_data

---