Juewen Fan
Gradient-Plateau SBA
Overall mean ASR: 0.962875
Highest average across all 8 settings.
Competition Overview
Artificial intelligence (AI), represented by deep learning, has deeply penetrated multiple fields, including production, daily life, and scientific research, and has become one of the core technological engines driving the progress of human society. AI replaces repetitive labor through automation and continues to push the boundaries of human cognition and problem-solving capabilities.
However, substantial research has demonstrated that AI models face serious security and privacy threats, including adversarial examples. Adversarial examples - carefully crafted inputs that mislead AI systems - pose significant threats to the reliability and safety of these models. By investigating adversarial examples and developing efficient attack-testing methodologies, we can not only assess model security but also enhance the intrinsic trustworthiness of AI systems.
This competition studies adversarial-example generation for multi-label classifiers in a strict black-box setting, where an attack may query only the model inputs and outputs. The challenge focuses on random hybrid attacks: for each image, one originally positive label must be hidden and one originally negative label must be added while all remaining labels are preserved.
Competition Results
Final ranking rule: the competition's final decision is to rank competitors by the arithmetic mean of their Mean ASR values across all eight model-dataset settings. The highest overall mean ASR wins. All scores are taken exclusively from the latest files in the official participant-information folder.
Chaofan Xie
Multi-Archive Search with SBA
Overall mean ASR: 0.823625
Second-highest average across all 8 settings.
Gaoren Zhang
Structured Active-Label Attack
Overall mean ASR: 0.763125
Third-highest average across all 8 settings.
| Model | Dataset | Juewen Fan | Gaoren Zhang | Chaofan Xie |
|---|---|---|---|---|
| ML-GCN | VOC 2007 | 0.999 (1) | 0.977 (2) | 0.933 (3) |
| ML-GCN | VOC 2012 | 0.996 (1) | 0.978 (2) | 0.943 (3) |
| ML-GCN | NUS-WIDE | 0.913 (1) | 0.020 (3) | 0.665 (2) |
| ML-GCN | COCO | 0.984 (1) | 0.890 (2) | 0.822 (3) |
| ML-LIW | VOC 2007 | 0.992 (1) | 0.918 (3) | 0.925 (2) |
| ML-LIW | VOC 2012 | 0.995 (1) | 0.957 (2) | 0.908 (3) |
| ML-LIW | NUS-WIDE | 0.944 (1) | 0.718 (2) | 0.715 (3) |
| ML-LIW | COCO | 0.880 (1) | 0.647 (3) | 0.678 (2) |
| Overall mean ASR / final place | 0.962875 / 1st | 0.763125 / 3rd | 0.823625 / 2nd | |
Values are the submitted Mean ASR followed by the per-setting rank in parentheses. All eight settings use Mean ASR; no Best-ASR proxy values are used.
Evaluation Protocol
- Models: ML-GCN and ML-LIW.
- Datasets: PASCAL VOC 2007, PASCAL VOC 2012, MS-COCO, and NUS-WIDE.
- Input size: 448 × 448 × 3.
- Perturbation bound: L2 norm no greater than 77.596.
- Query budget: no more than 10,000 fitness evaluations per image.
- Primary metric: mean attack success rate (ASR) over fixed runs.
- Overall ranking: arithmetic mean of the eight Mean ASR values; the highest overall mean ranks first.
Resources
Benchmark source code: CEC2026_MLA-EC on GitHub
Technical report: Benchmark for the CEC 2026 MLA-EC Competition
Results materials: every published score is taken from the latest participant materials in the official competition archive: Juewen Fan's report, Gaoren Zhang's revised results table, and Chaofan Xie's report.
Organizers
- Wenjian Luo — Harbin Institute of Technology, Shenzhen — luowenjian@hit.edu.cn
- Yiya Diao — Harbin Institute of Technology, Shenzhen — diaoyiya@hit.edu.cn
- Zhijian Chen — Harbin Institute of Technology, Shenzhen — 21B951010@stu.hit.edu.cn
- Yuhui Shi — Southern University of Science and Technology — shiyh@sustech.edu.cn