Applied ML researcher (UMBC) working across numerical optimization, compression-based learning, and ML for cybersecurity.
5
papers here
5
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
No publication-volume data is available in this snapshot, so output is unscored. This axis would measure total papers per year, authorship roles, and career-wide productivity trends across venues.
Impact—
Impact cannot be assessed: no citation counts, h-index, or downstream usage metrics were provided. This axis would measure citations, method adoption, and influence on subsequent research.
Focus0/100
Focus scores 0 because the snapshot shows an almost perfectly even spread: five areas — cybersecurity/applied ML (SaaS abuse and anomaly detection), software security/program analysis, machine learning/numerical optimization, ML for cybersecurity/software engineering, and compression-based ML (information theory) — each account for exactly 0.2 of the field distribution. This breadth fits a methods-plus-security portfolio, but no single subfield yet dominates the cataloged record.
Collaboration55/100
Collaboration scores 55. The snapshot records 15 distinct coauthors across the five cataloged papers, with recurring partners such as Michael Brautbar, Chang Liu, Kristopher Micinski, James Holt, and Vitaly Zaytsev. Authorship positions rotate across first, middle, and last, indicating both leading and supporting roles on moderately sized teams that span academic machine learning and industry security research, including the CrowdStrike abuse-detection study.
Activity—
No temporal publication cadence is available beyond the five 2026 arXiv entries, so activity is unscored. This axis would measure output per year and submission rhythm over the research career.
Reproducibility48/100
Reproducibility scores 48, averaged over 5 rated papers. Evolve rates four papers medium readiness and one — the deployed CrowdStrike abuse-detection system — low, reflecting proprietary data and limited methodological detail. Recurring review critiques, such as missing significance testing in the compression survey and confounded timing without ablations in the file-type study, indicate strong engineering results that stop short of fully reproducible evaluation.
Research summary
Edward Raff (University of Maryland, Baltimore County; ORCID 0000-0002-9900-1972) is an applied machine learning researcher whose cataloged 2026 output spans numerical optimization, compression-based learning, and machine learning for cybersecurity and software analysis. Evolve's catalog holds five arXiv papers, all from 2026. As first author, Raff's benchmarking of L1-penalized logistic regression shows that older solvers such as Liblinear and simple L-BFGS variants outperform the literature at practical penalty values. A co-authored survey formalizes lossless-compression methods — Normalized Compression Distance and Minimum Description Length — into a four-choice design framework, reporting competitive results that are notably strong on malware. As last author, a file-type identification study reports its LABELER model beating Magika with Macro F1 0.98 versus 0.91 across 54 text file types at 3.7x the speed, though Evolve flags confounded gains and missing ablations. Further collaborations cover a deployed CrowdStrike abuse-detection system built on Isolation Forest and expert risk scores (35% coverage, 30-58% fewer alerts), plus a fuzzing-based evaluation showing LLM decompilers can recompile more while preserving less behavior. The snapshot records 15 distinct coauthors and an even split of 0.2 across five fields, producing a focus score of 0; collaboration scores 55 and reproducibility 48 across five rated papers, while output, impact, and activity lack sufficient data. The profile is name-resolved from public bibliographic data.
Collaboration pattern
Raff's cataloged work is consistently team-based: five 2026 papers draw on a pool of 15 distinct coauthors, with repeated partners including Michael Brautbar, Chang Liu, Kristopher Micinski, James Holt, and Vitaly Zaytsev. Authorship rotates across first, middle, and last positions, and the portfolio mixes academic machine learning with industry security research, most visibly the CrowdStrike abuse-detection study. The collaboration score of 55 reflects an active but moderately sized network.
Research keywords
Highlights
Old, simple L1 solvers win at practical penalty settings
A design framework for compression-based machine learning
LABELER claimed to beat Magika on text file type identification
Passing shipped tests can mask LLM decompiler behavior drift
Deployed ML for security-platform product abuse detection
Timeline
- 2026First-author benchmarking of L1-penalized logistic solvers shows simple Liblinear and L-BFGS baselines beating the literature at practical penalty values.
- 2026Co-authors a survey formalizing compression-based machine learning (Normalized Compression Distance, Minimum Description Length) as a four-choice design framework.
- 2026Last-author work introduces LABELER for text-content file type identification, reporting Macro F1 0.98 versus Magika's 0.91 at 3.7x the speed.
- 2026Contributes to a CrowdStrike study on a deployed ML abuse-detection system for security platforms, combining Isolation Forest with expert risk scores.
- 2026Co-authors a fuzzing-based differential evaluation showing LLM decompilers can recompile more successfully while preserving less behavior.
Papers in the Evolve catalog
- first author
Simpler Methods Work Better for L1 Penalized Logistic Models and Large Datasets
An Introduction to Compression-Based Machine Learning
Fast And Accurate Text Content File Type Identification
Identifying Security Platform Product Abuse with Machine Learning
When LLM Decompilers Recompile More and Preserve Less
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.
- Dimension scores come from a bibliometric snapshot only; no citation counts, h-index, or venue-level metrics were provided.
- All five cataloged papers are 2026 arXiv preprints; peer-reviewed venues and pre-2026 career output fall outside this profile's data.
- Output, impact, and activity axes lack sufficient underlying data and are intentionally left unscored.
- Profile is name-resolved from public bibliographic data tied to ORCID 0000-0002-9900-1972; 'Edward Raff' is a shared name and identity beyond this record is not independently verified.