FigureAsia 35 Under 35 · Science
Yiming Li
Age 29 · AI security and data rights · China / Singapore
AI-security researcher turning model ownership into a testable technical claim without planting a label-changing backdoor.
- Approximate age at the edition eligibility date
- 29
- Field
- Computer science
- Country or region
- China / Singapore
- FigureAsia U35 Assessment
- 80.6 / 100
Profile
Career and documented record
Yiming Li works on a question with immediate commercial consequences: how a developer can show that a neural network was copied without planting a backdoor in the original model. In June 2025, IEEE Transactions on Pattern Analysis and Machine Intelligence published MOVE, with Li first on the eight-author paper, in volume 47, issue 6, pages 4734–4751.
MOVE applies style transfer to a small share of training images while retaining their labels, then trains a meta-classifier to test whether a suspect model carries those defender-selected external features. The journal version extends the team's 2022 AAAI work from white-box to black-box verification and adds adaptive-attack analysis and a broader experimental record.
Li is now a Research Fellow at Nanyang Technological University, where he frames his programme around responsible and sustainable AI, AI risk management and copyright protection. MOVE provides evidence under tested image-classification threat models; it is neither a legal finding of infringement nor a universal guarantee against model theft.
FigureAsia selection
Why Yiming Li is on the list
Li is selected for turning model ownership from a policy assertion into a measurable security problem. MOVE addresses copying without modifying class labels, and its first authorship makes his responsibility legible.
Verified work
The 2025–26 record
MOVE in IEEE TPAMI
First-listed author of the eight-author journal paper in volume 47, issue 6, pages 4734–4751; DOI 10.1109/TPAMI.2025.3546223.
Ownership verification without a backdoor
Embedded defender-selected external features through label-preserving style transfer and tested for their presence with a meta-classifier.
Expanded threat model
Extended the earlier AAAI study to white-box and black-box settings, with adaptive-attack analysis and broader experiments.
Field context
The work in its field
Neural-network ownership is difficult to prove because weights can be hidden, models can be modified and conventional backdoor watermarks may compromise normal behaviour. A useful method must survive realistic access limits without creating a new security risk.
FigureAsia U35 Assessment
Assessment breakdown
80.6out of 100
Substantive 2025–2026 contribution
15 / 20
First-listed author of the eight-author journal paper in volume 47, issue 6, pages 4734–4751; DOI 10.1109/TPAMI.2025.3546223.
Verified scientific impact
11.3 / 15
A first-authored TPAMI result turns AI ownership from a policy assertion into a measurable security problem while avoiding the label-changing backdoors used by many earlier watermarking schemes.
Originality and distinction
8.2 / 10
The distinction lies in using label-preserving style transfer and a meta-classifier to verify copied-model ownership without inserting a conventional backdoor.
Field influence
8.3 / 10
Researchers in ai security and data rights now have a stronger result to test, extend or challenge because of this contribution.
Individual agency
8.2 / 10
Li is the first-listed author of the eight-author TPAMI paper. The evidence supports lead authorship of the journal extension, not sole invention.
Durability and trajectory
4.3 / 5
MOVE extends Li's earlier AAAI work into a fuller journal record and sits inside his continuing NTU programme on AI risk and copyright protection.
Asian significance and global relevance
4.3 / 5
Chinese computer scientist educated at Tsinghua and Zhejiang Universities and now working in Singapore.
Evidential validity and reproducibility
6.7 / 8
The paper tests white-box and black-box settings and adaptive attacks; no claim of universal robustness or legal proof is made.
Advance in scientific knowledge
5.8 / 7
MOVE shows how externally selected visual features can carry a detectable ownership signal without changing training labels.
Translational or methodological utility
4.3 / 5
The method gives developers a technical route to investigate suspected model copying when weights or training records are unavailable.
Responsible research stewardship
4.2 / 5
The profile separates experimental ownership evidence from legal judgement and retains the limits of the tested threat models.