Md Khalid Syfullah
Bangladesh Army University of Science and Technology
Privacy-preserving split & federated learning researcher at Bangladesh Army University of Science and Technology
3
papers here
3
reviewed by Evolve
3
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 scoreable: the cataloged work consists of three 2026 arXiv preprints and OpenAlex provides no output indicators beyond them, so this axis — which would measure publication volume, venue quality, and preprint-to-published conversion — cannot be estimated.
Impact—
Citation impact cannot be scored yet: h-index and i10-index both read 0 on OpenAlex because the 2026 preprints are too new for citations to accumulate, so this axis — measuring citation uptake, influence on follow-up work, and venue visibility — is left unscored.
Focus0/100
Focus scores 0 because output is spread evenly rather than concentrated: the field distribution splits 0.33/0.33/0.33 across privacy-preserving distributed machine learning for medical VQA, machine learning for mental health, and trustworthy computer vision. Topic data confirms the breadth — privacy-preserving technologies (3 papers) leads, but neural network applications, handwritten text recognition, multimodal learning, and educational technology (2 each) also appear, so split learning is the only throughline.
Collaboration14/100
Collaboration scores 14, the weakest present axis: OpenAlex records only 1 distinct coauthor, Alvi Ataur Khalil, who appears as the recurring partner across the 2026 preprints. All three cataloged papers are first-authored and none lists the researcher as corresponding author, suggesting a small two-person working mode rather than lab-scale or consortium research. Broadening the co-author network and taking corresponding responsibility would strengthen this dimension.
Activity61/100
Activity scores 61 on strong recent momentum: OpenAlex counts 7 works in the last three years, all concentrated in 2026, the sole year appearing in countsByYear. Three of these are first-authored arXiv preprints reviewed by Evolve, indicating a burst-style cadence typical of an early-career researcher establishing a niche; sustaining output beyond 2026 will determine whether this pace holds.
Reproducibility43/100
Reproducibility scores 43, the mean readiness across 3 reviewed papers. Evolve rates the split-learning mental-health paper low (lenient ANLS metric, privacy claim limited to data locality) and the other two medium; recurring flags include single-run experiments, IID-only USPLIT-VQA results, and low attack headroom in the token-pruning evaluation. Releasing code, multi-seed results, and stronger threat-model baselines would raise this score materially.
Research summary
Md Khalid Syfullah is an early-career researcher at Bangladesh Army University of Science and Technology working at the intersection of privacy-preserving distributed machine learning and multimodal AI. The 2026 catalog shows three first-authored arXiv preprints, all co-authored with Alvi Ataur Khalil, that form a coherent research program: adapting split learning — and its U-shaped variant — so that vision-language models and LLMs can be trained across parties without sharing raw data. USPLIT-VQA introduces contribution-aware weighted aggregation with Byzantine-robust reputation for medical visual question answering; DTP reframes question-guided token pruning as a privacy defense for distributed VLM VQA; and a schema-aware split-learning framework fine-tunes a LoRA-adapted LLaMA-3.2-3B for mental distress prediction across heterogeneous surveys, reporting 0.708 ANLS and beating FedAvg in 8 of 9 settings. OpenAlex attributes 7 works in the last three years, all in 2026, with h-index and i10-index still at 0 — consistent with a brand-new output record. Evolve's reviews are constructive but flag methodological caveats: single-run experiments, IID-only evaluation, low attack headroom, and privacy guarantees limited to data locality. Reproducibility readiness averages 43/100 across the three papers. The profile is name-resolved from OpenAlex bibliographic data; no ORCID is linked.
Collaboration pattern
The record shows an unusually tight two-person pattern: OpenAlex lists only 1 distinct coauthor, and Alvi Ataur Khalil is the recurring partner across the 2026 preprints. Syfullah is first author on all three cataloged papers while none lists him as corresponding author, and the collaboration dimension scores 14/100. Independent verification of contribution norms and a wider network would clarify how this small-team output is produced.
Research keywords
Highlights
Schema-aware split learning tunes LLaMA-3.2-3B for mental distress prediction
Question-guided token pruning as a privacy defense for vision-language models
U-shaped split learning for VQA with reputation-based robust aggregation
Timeline
- 2026Released USPLIT-VQA, a U-shaped split-learning framework for visual question answering with contribution-aware weighted aggregation and Byzantine-robust reputation.
- 2026Proposed question-guided token pruning (DTP) as a privacy defense for vision-language models in distributed VQA.
- 2026Published LLM-based schema-aware split learning for mental distress prediction, reporting 0.708 ANLS and beating FedAvg in 8 of 9 settings.
- 2026Accumulated 7 works within 2026 per OpenAlex countsByYear, with 3 first-authored preprints entering the Evolve catalog.
Papers in the Evolve catalog
- first author
LLM-Based Schema-Aware Split Learning for Privacy-Preserving Mental Distress Prediction Across Heterogeneous Surveys
- first author
Don't Send What You Don't Need: Question-Guided Token Pruning as a Privacy Defense for Vision-Language Models
- first author
USPLIT-VQA: U-Shaped Split Learning for Visual Question Answering with Contribution-Aware Weighted Aggregation
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 OpenAlex only (identity: openalex); no ORCID is linked and the name is common, so this profile is name-resolved from public bibliographic data.
- All cataloged outputs are 2026 arXiv preprints; no peer-reviewed venue, citation, or funding information is available yet.
- Output and impact dimensions are unscored: h-index and i10-index both read 0 and citation data have not accumulated.
- Reproducibility evidence is limited to Evolve's review of 3 papers, which flag single-run experiments and low-to-medium readiness.