Mechanistic interpretability and federated-learning privacy researcher at the University of the West of England
2
papers here
2
reviewed by Evolve
1
first-author
0
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—
Insufficient data: no publication-count or venue records are available for this profile yet. This axis would measure total research output volume, the distribution across venues and preprint servers, and the share of first- or corresponding-author papers.
Impact—
Insufficient data: no citation counts, h-index or influence metrics exist in the input data. This axis would measure citation impact and the uptake of the reviewed arXiv preprints by other researchers.
Focus0/100
The focus score of 0 reflects maximal dispersion, not inactivity: the field distribution across the two reviewed papers is an exact 50/50 split between machine learning security and privacy (federated learning) and machine learning — mechanistic interpretability. With only two reviewed papers, one in each field, no dominant direction has emerged yet; a larger sample would show whether this is a deliberate dual-track agenda or early-stage exploration.
Collaboration46/100
The collaboration score of 46 sits mid-range and is computed from 9 distinct coauthors across the reviewed papers. The profile shows team-based work rather than solo output: Muhammad Khan is first author on the interpretability paper and a middle author on the federated learning paper, neither as corresponding author. Recurring collaborators include Ali Akarma, Toqeer Ali Syed, Qurat-ul-ain Mastoi, Adeel Ahmad, Shlok Channawar, Akshaj Gurugubelli, Girish Gupta and Aditya Shah.
Activity—
Insufficient data: only 2026-dated papers are available, too narrow a window to compute a publication cadence. This axis would measure publishing frequency, year-over-year continuity, and sustained research activity across recent years.
Reproducibility55/100
Reproducibility scores a mean of 55 across the 2 papers Evolve has reviewed, both rated at medium reproducibility readiness. This is a moderate signal: the steerability study is described as modest but well-controlled, though its evidence is rank-only (ρ reaching −0.546), and the privacy study is solid in its clip-then-noise defense analysis but carries an internally inconsistent DP budget table. Neither paper has independent replications in this dataset.
Research summary
Muhammad Khan is a researcher at the University of the West of England whose reviewed work spans two areas of machine learning in equal measure: mechanistic interpretability and the security and privacy of federated learning. The Evolve catalog holds two 2026 arXiv papers. The first-authored “Look Before You Steer: Geometry Predicts SAE Feature Steerability” proposes decoder-geometry metrics that rank how steerable sparse-autoencoder features are before any intervention; Evolve judged the evidence modest but well-controlled, with rank-only correlations (ρ up to −0.546), credible with caveats. The co-authored privacy paper reports near-total (~100%) client de-anonymization from gradient updates on a human-activity-recognition proxy in vehicular edge networks, alongside a clip-then-noise differential-privacy defense study whose DP budget table Evolve flagged as internally inconsistent. The radar shows a mid-range collaboration score (46, from 9 distinct coauthors), a reproducibility mean of 55 across the two reviewed papers (both medium readiness), and a focus score of 0, driven by the exact 50/50 field split between the two topics — an early-stage, dual-track portfolio rather than an established single-line agenda. Output, impact and activity axes lack sufficient bibliometric data, so no citation-based assessment is possible at this time.
Collaboration pattern
Work is team-based: 9 distinct coauthors appear across the two reviewed papers, giving a mid-range collaboration score of 46. Muhammad Khan leads as first author on the interpretability paper and contributes as a middle author on the federated learning paper, neither as corresponding author. Recurring collaborator names include Ali Akarma, Toqeer Ali Syed, Qurat-ul-ain Mastoi, Adeel Ahmad, Shlok Channawar, Akshaj Gurugubelli, Girish Gupta and Aditya Shah, spanning the reviewed 2026 output.
Research keywords
Highlights
Predicting SAE feature steerability from decoder geometry
~100% client de-anonymization from federated learning updates
A balanced dual-track portfolio at an early stage
The field distribution across the two reviewed papers is an exact 50/50 split between federated-learning privacy and mechanistic interpretability, with a focus score of 0 and a reproducibility mean of 55 (both papers medium readiness) — an early-stage profile spanning model internals and model privacy.
Timeline
- 2026First-authored “Look Before You Steer: Geometry Predicts SAE Feature Steerability” released on arXiv, opening the mechanistic-interpretability line of work.
- 2026Co-authored the federated learning privacy study on gradient-based client identity inference and defenses for inertial sensing in vehicular edge networks, released on arXiv.
- 2026Both papers entered the Evolve catalog with medium reproducibility readiness, yielding a reproducibility mean of 55 and an exact 50/50 field split across the two topics.
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 scores derive from OpenAlex-based public data; citation counts, h-index and publication totals were unavailable, so the output, impact and activity radar axes are marked missing.
- “Muhammad Khan” is a common name and no ORCID is linked; this profile is name-resolved from public bibliographic data and could merge or split distinct individuals.
- Evidence rests on only two 2026 arXiv preprints reviewed by Evolve; peer-review status, venue acceptance and independent replications are not reflected in this profile.
- Both Evolve verdicts carry methodological caveats — rank-only correlations (ρ up to −0.546) in the steerability study and an internally inconsistent DP budget table in the privacy study.