Guangdong Bai

City University of Hong Kong · China

Unclaimed · AI-structured profileORCID OpenAlex

AI security researcher bridging adversarial robustness, private retrieval, and molecular graph generation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

The output axis is unscored in this snapshot: the available records provide partial counts but no consolidated total-volume measure. This axis would capture cumulative publication count and its growth across the career; Evolve will add a score only once the full work record is indexed.

Impact—

Impact is not yet scored. Raw indicators do exist — h-index 23 and i10-index 55 — but citation totals and field-normalized influence have not been computed. This axis would measure citation impact relative to field norms and career stage, and remains deliberately unscored here.

Focus0/100

The focus score of 0 reflects an unusually broad portfolio: the field distribution is split exactly evenly at 0.5 between AI security / privacy-preserving information retrieval and machine learning for molecular graph generation, and output spans five topic clusters — adversarial robustness (69), privacy-preserving technologies (54), security and verification (53), privacy and data protection (30), and malware detection (30). The 2026 preprints confirm this dual-track pattern.

Collaboration54/100

Collaboration scores 54 on the 0–100 scale. The recent window records 14 distinct coauthors, and 12 recurring collaborators are listed, including Xinguo Feng, Zihan Wang, Kaiyan Wen, Shijie Zhang, Lu Yu, and Alsharif Abuadbba. On all three 2026 catalog papers Bai holds the last author position without the corresponding-author flag, consistent with senior supervision of small-to-mid-sized teams rather than large consortia.

Activity100/100

Activity is the standout axis at 100/100. The counts-by-year series climbs from 17 works (2022) to 22 (2023), 33 (2024), a peak of 43 (2025), and 39 (2026), while the recent three-year window contains 115 works — roughly 38 per year. This sustained, accelerating trajectory places the researcher in the platform's most active band.

Reproducibility55/100

Reproducibility scores 55, computed from 2 assessed papers with a mean readiness of 55; all three catalog preprints carry a medium repro-readiness rating. Evolve's reviews flag concrete gaps — EGF's learned-geometry claim lacks an ablation, SHAQ is tested only on 32-token documents, and CRACK shows thin statistics with unproven surrogate transfer — while each paper reports checkable metrics.

Research summary

Guangdong Bai, based at City University of Hong Kong, holds a full-tier OpenAlex-resolved profile (ORCID 0000-0002-6390-9890) at the intersection of AI security and generative machine learning. The field split is exactly even — 0.5 for AI security / privacy-preserving information retrieval and 0.5 for machine learning with generative models for molecular graphs — and the leading topics confirm a security-first core now extended into generative modeling: adversarial robustness (69), privacy-preserving technologies (54), and security and verification (53). Raw indicators include an h-index of 23, an i10-index of 55, and 115 works in the recent three-year window, with annual counts climbing from 17 in 2022 to 22 (2023), 33 (2024), a peak of 43 (2025), and 39 (2026) — the basis for a maximal activity score of 100. In 2026, three arXiv preprints with Bai as last author entered the Evolve catalog: CRACK, which jailbreaks layered text-to-image defenses with up to 99.63% attack success in as few as 6 queries; SHAQ, which cuts vec2text embedding-inversion recovery to roughly 0.27 R-1 at near-baseline retrieval utility; and EGF, which beats DiGress and GruM on all four QM9 metrics. All three carry a medium repro-readiness rating, consistent with a reproducibility score of 55.

Collaboration pattern

Bai works within a stable, mid-sized network: 14 distinct coauthors in the recent window and 12 recurring collaborators, including Xinguo Feng, Zihan Wang, Kaiyan Wen, Shijie Zhang, Lu Yu, Alsharif Abuadbba, and Chris Siu Yeung Chow. On all three 2026 catalog papers he takes the last author slot without the corresponding-author flag, suggesting a supervisory role in which junior colleagues drive experiments across both the security and generative-modeling tracks.

Research keywords

adversarial robustnessAI safetyjailbreak attackstext-to-image generationmulti-agent debateprivacy-preserving retrievalembedding inversion attacksvector databasesretrieval-augmented generationgraph generationflow matchingmolecular graph generation

Highlights

Multi-agent debate cracks layered text-to-image safety filters
The CRACK preprint reports jailbreaking heterogeneous, multi-layer text-to-image defenses with up to 99.63% attack success rate in as few as 6 queries, using black-box adversarial search with multi-agent debate and reinforcement learning. Evolve calls the results compelling but flags unproven surrogate transfer and thin statistics.
Shadow queries blunt embedding inversion in vector databases
SHAQ replaces stored document embeddings with shadow-query embeddings, cutting vec2text recovery to roughly 0.27 R-1 while keeping dense-retrieval utility near baseline; Evolve notes that evaluation is limited to 32-token documents.
Embedded graph flows advance categorical molecular generation
EGF learns endpoint anchors for categorical graph generation and beats DiGress and GruM on all four QM9 metrics, combining flow matching with learned category embeddings and permutation equivariance; Evolve credits the gains but notes the learned-geometry claim still lacks an ablation.

Timeline

2010Earliest year in the provided counts-by-year series, with 2 works recorded.
2022The annual count turns upward at 17 works, up from 5 in 2019, beginning a sustained climb.
202322 works recorded in the year counts as the ramp continues.
202433 works recorded, the third consecutive annual increase in the series.
202543 works — the peak of the provided counts-by-year series.
202639 works recorded; three arXiv preprints enter the Evolve catalog — EGF, SHAQ, and CRACK — each with Bai in the last author position.

Papers in the Evolve catalog

Frequent co-authors

Xinguo Feng×2 Zihan Wang×2 Kaiyan Wen×1 Shijie Zhang×1 Lu Yu×1 Zhongkui Ma×1 Chuan Yan×1 Guowei Yang×1 Alsharif Abuadbba×1 Ethan Ma×1 Chris Siu Yeung Chow×1 Qingqing Li×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometrics come from an OpenAlex snapshot only (identity: openalex); no Scopus, Web of Science, or DBLP cross-check is included.
The output and impact axes are unscored in this snapshot; the h-index of 23 and i10-index of 55 are raw indicators, not normalized impact measures.
Only three 2026 arXiv preprints from the Evolve catalog are individually reviewed; peer-review status and published versions are not reflected.
The profile is name-resolved from public bibliographic data anchored to ORCID 0000-0002-6390-9890; 'Guangdong Bai' is a shared name and its Chinese character form is unverified.