Adversarial robustness researcher: label-free auditing of LLM utility degradation and wearable activity recognition
2
papers here
0
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-27) 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—
Insufficient data: the snapshot has no aggregated publication counts or venue records, so output cannot be scored. This axis would measure the volume and consistency of publications over time across venues and years.
Impact—
Insufficient data: no citation counts, h-index, or downstream-uptake records are available, so impact cannot be scored. This axis would measure how often and how widely the work is cited and built upon by other researchers.
Focus—
Insufficient data: with only two cataloged 2026 papers, topical concentration cannot be quantified. This axis would measure how consistently the research agenda stays concentrated within and across fields over multiple years.
Collaboration39/100
Scores 39 of 100, computed from bibliometric data recording 6 distinct coauthors in the current snapshot: Qingyu Wu, Zeyu Feng, Yongda Yu, Yuzhe Luo, Yuan Wei, and Renju Liu. Both 2026 cataloged papers place Cheng in the last author position, consistent with leading a small, stable group; the mid-range score and compact coauthor pool suggest a focused network rather than a broad multi-team one.
Activity—
Insufficient data: the snapshot provides no publication-frequency timeline beyond the two 2026 arXiv entries, so activity cannot be scored. This axis would measure publication cadence and sustained research engagement in recent years.
Reproducibility—
Insufficient data: no computed reproducibility score exists yet, though Evolve tags both cataloged papers as medium reproducibility readiness. This axis would measure artifact availability, code release, and how reliably experiments can be replicated.
Research summary
Hua Cheng is a machine learning researcher whose cataloged work centers on adversarial robustness and AI safety. The Evolve catalog holds two 2026 arXiv papers, in both of which Cheng is the last-listed author (not the marked corresponding author) — a placement usually associated with senior or supervisory responsibility. The first line of work, ENDOPROMPT, targets LLM security: it learns utility-degrading prefixes from unlabeled victim signals, reporting a -26.8 percentage-point effect across 27 of 28 evaluation cells, built on pseudo-references, preference optimization, and GRPO; Evolve's verdict calls the work honest while noting its request-matching advantage remains unproven. The second paper applies a label-sealed select-then-project audit to wearable activity recognizers, finding that universal adversarial perturbation failures are mostly DC-carried; Evolve describes it as rigorous and honest, with a caveat about possible circularity in DC-retained search. The bibliometric footprint is compact: 6 distinct coauthors and a collaboration dimension score of 39, while the output, impact, focus, activity, and reproducibility axes lack sufficient data to score. Both papers are rated medium for reproducibility readiness. This is an early-stage profile built from a narrow snapshot rather than a full publication record.
Collaboration pattern
Hua Cheng's cataloged work shows a compact collaboration footprint: 6 distinct coauthors in the current snapshot — Qingyu Wu, Zeyu Feng, Yongda Yu, Yuzhe Luo, Yuan Wei, and Renju Liu — yielding a collaboration dimension score of 39. In both 2026 papers Cheng sits in the last author position without being the marked corresponding author, a pattern consistent with supervising or advising a small, stable team rather than operating inside a large multi-institution consortium.
Research keywords
Highlights
ENDOPROMPT: victim-side pseudo-references degrade LLM utility
Label-sealed audit of wearable activity recognizers
A label-free throughline across two robustness domains
Both 2026 papers share one methodological stance — deriving attack or audit signals without labels: unlabeled victim signals for LLM utility degradation, and a label-sealed audit for wearable recognizers — with Cheng in the last author position on each.
Timeline
- 2026ENDOPROMPT appears on arXiv, introducing victim-side pseudo-references for LLM utility degradation; Cheng is last author.
- 2026The label-free auditing study of wearable activity recognizers appears on arXiv; Cheng is again last author.
- 2026The two papers together anchor an adversarial-robustness profile spanning LLM security and wearable sensing, with 6 distinct coauthors in the snapshot.
Papers in the Evolve catalog
Frequent co-authors
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.
- Bibliometric signals come from a sparse snapshot: no institution, country, or ORCID is attached, and the output, impact, focus, activity, and reproducibility axes have insufficient underlying data.
- "Hua Cheng" is name-resolved from public bibliographic data; without an ORCID or a verified Chinese-character name form, overlap with other same-name researchers cannot be ruled out.
- Both cataloged papers are 2026 arXiv preprints; peer-review status, venue acceptance, and citation history are not reflected in this profile.
- Only statistics supplied in the Evolve catalog and radar snapshot are quoted (e.g., -26.8pp across 27/28 cells, 6 coauthors, collaboration score 39); no external citation or publication database was consulted.