Fnu Suya

The University of Texas at Austin · United States

Unclaimed · AI-structured profileORCID OpenAlex

ML security and systems researcher at UT Austin studying LLM inference reliability and privacy attacks on speech models

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—

Not yet scored. This axis measures publication volume and productivity over time. The snapshot's counts by year hint at accelerating output (9 works in 2026 versus 2 in 2025), but consolidated output data is insufficient to score it.

Impact—

Not yet scored. This axis measures citation-based influence such as citation counts and h-index trajectory. The snapshot reports an h-index of 7 and an i10-index of 5, but citation-level detail is insufficient to compute the impact score.

Focus100/100

Focus scores 100, the strongest signal in this profile. The topic distribution is tightly concentrated: adversarial robustness in machine learning leads with 22 records, followed by security and verification in computing (15) and privacy-preserving technologies (13). The reviewed catalog clusters into a single declared field, machine learning systems for LLM inference, consistent with a security-and-reliability lens applied across model types.

Collaboration51/100

Collaboration scores 51, a mid-range reading. The snapshot lists 12 distinct coauthors, with frequent partners including Abeer Alwan, Yuan Tian, Deepayan Chakrabarti and Xueping Li. On both reviewed 2026 papers Suya appears in a middle author position without corresponding status, suggesting contribution within larger team projects rather than solo or lead authorship so far in the catalog.

Activity75/100

Activity scores 75. The counts-by-year series shows sustained output with a pronounced recent acceleration: after sparse early years (1 in 2017, 2 in 2019, 3 in 2020) and a 2023 peak of 7 works, 2026 already registers 9 works, and 12 works fall within the most recent three years, indicating a strongly active current phase.

Reproducibility100/100

Reproducibility scores 100, though it rests on a single evaluated paper (reproMean 100 from 1 paper). That paper — the cross-precision LLM inference study — is rated high, with well-controlled experiments and honestly bounded claims; the TTS membership-inference paper is rated medium, with a single-split evaluation noted as limiting certainty.

Research summary

Fnu Suya is a machine learning security and systems researcher at The University of Texas at Austin whose work, per an OpenAlex-resolved bibliometric snapshot (ORCID 0000-0002-3334-6257), concentrates on attacking and hardening learned models. The snapshot's topic profile is led by adversarial robustness in machine learning (22 records), followed by security and verification in computing (15) and privacy-preserving technologies (13), with smaller footprints in digital media forensics (6) and ML-based classification (3). Output is accelerating sharply: after sparse early years (1 work in 2017, 2 in 2019, 3 in 2020), the counts-by-year series peaks at 7 works in 2023 and already registers 9 in 2026, with 12 works in the most recent three years. The reported h-index is 7 with an i10-index of 5. Evolve has reviewed two 2026 arXiv papers, both with Suya in a middle author position. The first shows that greedy decoding is not precision-invariant — BF16 and FP16 inference diverge on 49–100% of prompts — and proposes a margin-gated FP32 lm_head fix recovering 22–36 percentage points on an A10G GPU; it rates high for reproducibility. The second reports the first black-box membership inference attacks on fine-tuned text-to-speech models, reaching AUC above 0.8 with careful ablations, though a single-split evaluation limits certainty. Twelve distinct coauthors appear in the record, and the collaboration dimension scores 51, consistent with steady team-based work.

Collaboration pattern

Twelve distinct coauthors appear in the snapshot, with frequent collaborators including Kunlin Cai, Kaiyuan Zhang, Zihang Xiang, Jinghuai Zhang, Abeer Alwan, Yuan Tian, Gaoyuan Du, Anam Nawaz Khan, Rex Zhou, Xiaoyang Liu, Deepayan Chakrabarti and Xueping Li. Both reviewed 2026 papers place Suya in a middle author position without corresponding status, and the collaboration dimension scores 51 — consistent with steady team-based work rather than dominant senior or solo authorship.

Research keywords

Adversarial robustnessMembership inference attacksLLM inferenceNumerical precisionReproducibilityText-to-speechPrivacy auditingSecurity and verificationPrivacy-preserving machine learningDigital media forensicsVoice cloningMachine learning classification

Highlights

Greedy decoding is not precision-invariant in LLM inference
Evolve's review of this 2026 arXiv study finds BF16 and FP16 greedy decoding diverging on 49–100% of prompts; a margin-gated FP32 lm_head fix recovers 22–36 percentage points on an A10G GPU. Verdict: well-controlled, honestly bounded; reproducibility rated high.
First black-box membership inference attacks on fine-tuned TTS models
The 2026 arXiv study reports the first black-box membership inference attacks on fine-tuned text-to-speech models, achieving AUC above 0.8 with careful ablations; Evolve notes that a single-split evaluation without variance limits certainty. Reproducibility rated medium.
Research output accelerating into 2026
The snapshot's counts by year show 9 works recorded in 2026 versus 2 in 2025 and a previous peak of 7 in 2023, with 12 works in the most recent three years — the activity dimension scores 75.

Timeline

2017First year with recorded works in the OpenAlex snapshot: 1 work.
2020Early output builds up: 2 works recorded in 2019 and 3 in 2020 in the snapshot's counts by year.
2023Most productive earlier year in the snapshot, with 7 works recorded.
2025Output continues with 2 works recorded in the snapshot.
2026Sharpest activity yet: 9 works recorded, including the two Evolve-reviewed arXiv studies on cross-precision LLM inference divergence and black-box membership inference on fine-tuned TTS models.

Papers in the Evolve catalog

Frequent co-authors

Kunlin Cai×1 Kaiyuan Zhang×1 Zihang Xiang×1 Jinghuai Zhang×1 Abeer Alwan×1 Yuan Tian×1 Gaoyuan Du×1 Anam Nawaz Khan×1 Rex Zhou×1 Xiaoyang Liu×1 Deepayan Chakrabarti×1 Xueping 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.

Profile is name-resolved from public bibliographic data; identity confidence rests on the OpenAlex record and ORCID 0000-0002-3334-6257.
All bibliometrics (h-index 7, i10-index 5, counts by year, topics, 12 distinct coauthors) come from an OpenAlex snapshot and cover only indexed works.
The reproducibility score of 100 rests on a single evaluated paper; the output and impact radar axes are unscored because the underlying data is insufficient.
Evolve has reviewed only 2 of this researcher's papers, narrow coverage relative to the recorded output (e.g. 9 works in 2026).