Security engineering & adversarial machine learning. I build small, verifiable tools for detection work, and I try to be precise about what my numbers actually prove.
- S.Kom. / B.Comp.Sc. — Information Systems (Enterprise Systems), Institut Teknologi dan Bisnis STIKOM Bali. GPA 3.34/4.00, 144 credits (NIM 190030591). Thesis on campus-management optimisation treated as classification.
- M.Comp.Sc. (S2) in Cyber Security — starting February 2027.
- Based in Bali, Indonesia. Working in Indonesian and English.
My interest sits where detection engineering meets machine learning: what a model reports, and whether that report survives an adversary who knows the model exists.
Robustness of ML-based network intrusion detection (ML-NIDS) under problem-space evasion. Most adversarial-ML work on NIDS perturbs feature vectors; that is not what an attacker can actually do on a wire. My work evaluates detectors against evasions that remain valid at the protocol and traffic level, and reports the gap between feature-space and problem-space results.
Status (2026-11): four desk rejections, no peer review at any stage — IEEE TDSC (Sept 2026), JISA (Sept 2026, out of scope), Computers & Security (Sept 2026, AI/ML moratorium), Neurocomputing (Oct 2026, NEUCOM-D-26-23741). The retitled package for Expert Systems with Applications is assembled and not yet submitted. The preprint is still unposted.
Related threads: detection engineering (Suricata/Zeek rules, lab traffic capture) and reproducibility of ML experiments (fixed seeds, versioned data, one command to regenerate every table).
Kenapa Mesin Bisa Salah? / Why Machines Can Be Wrong?
— a tiered textbook in two parallel editions, Indonesian and English: 32 chapters in
six parts, 304 exercises with 304 answer keys, from linear algebra and probability to
adversarial machine learning, formal methods in Lean 4, and differential privacy. Each
chapter carries four reader paths (child, undergraduate, lecturer, researcher), and every
worked number in it comes from an actual execution rather than from memory. The book is
authored as LaTeX source, so a clean checkout rebuilds both PDFs: make check is the
quality gate (chapter count, glossary, exercise↔key mapping), make test runs 36
checks plus a sandbox compile, and a translation-parity gate keeps the two editions
from drifting. Text CC BY 4.0, code MIT.
paperlint — deterministic pre-submission
gates for manuscripts and research repositories, behind one interface, one exit code and
one JSON report so they run in CI instead of in someone's memory. Every check in it
started as a throwaway script written during a real submission round; between them they
caught three cited references that did not exist, a Word file shipping literal ?? where
cross-references should be, a figure overflowing its canvas by 91%, and a deposit record
marked "restricted" while claiming to be publicly archived. MIT.
rulebench — Suricata rule validator and
alert-coverage analyzer: is a ruleset internally consistent, and is it earning its keep?
suricata -T proves rules load; it says nothing about dead rules, colliding SIDs, or
alerts firing for signatures that no longer exist. rulebench answers those statically,
with no daemon, config file, or running Suricata instance. Rust, MIT.
anf-cyber-shell — zsh framework
for authorised security research. One dispatcher over 16 external security CLIs
(nmap, httpx, nuclei, ffuf, subfinder, amass, nikto, whatweb, gobuster, sqlmap-free
recon, whois, dig, yara, suricata, jq, curl), workspace-scoped artefacts, a JSONL
audit log, and GuardFall v2 — a command canonicaliser that rejects shell-injection
and destructive patterns before passthrough execution. 3,111 lines in a single zsh module, 116 subcommand functions; make verify
runs 15 checks (PASS=15 FAIL=0 locally; 14 pass + 1 skipped in CI, where one check
needs a tool the runner does not have). MIT.
security-lab — the CLI tooling I
actually use, consolidated: sift (zero-dependency Go secret scanner with git-history
walking), dockguard (Dockerfile linter), log-anonymizer, sec-scan-cli (npm audit /
pip-audit / gitleaks / Suricata in one pipeline), and pilgrims (modules for authorised
engagements) behind one seclab entry point. Secret-scan clean (gitleaks detect:
no leaks); the only credential-shaped strings in the repo are deliberately fake test
fixtures, allowlisted in .gitleaks.toml with the reason written next to them.
b-comp-sc-research —
undergraduate research on campus-management optimisation treated as a classification
problem across three domains (marketing → Random Forest, HR → Decision Tree,
operations → SVM/kNN). The README separates two kinds of numbers on purpose: the demo
pipeline is reproducible here and asserted by verify.py on the committed fixtures
(9–12 rows each, so a "100%" there means nothing about model quality), while the thesis
figures were computed on the full institutional dataset, which is not in the
repository and therefore cannot be regenerated from it.
Mini-SOC module for SMK (vocational high school), network security track — written, not yet delivered, so there are no student outcomes to report. It ships as a five-part package (module text, a 14-slide deck, student worksheet, rubric, and a lab) rather than slides alone. The lab is the part I care about: students get 248 synthetic log events with planted incidents, a detector skeleton that under-detects, and a verifier that fails if their detector over- or under-fires. The generator is seeded, so every run produces byte-identical logs and a student's "it works on my machine" is checkable.
Three rules I apply to my own output, including the repos above:
- Reproducibility over polish. If a number appears in a README or a paper, a script in the repo regenerates it from versioned inputs. A result I cannot re-run is a result I do not report. Every repo of mine ships a verification command, and I run it before I publish.
- No fabricated numbers. No placeholder metrics, no "approximately" F1 scores, no benchmark tables assembled from memory. When a claim stops being true, it is corrected in the README and the limitation is written down rather than quietly deleted.
- Backup before mutation. Full archive before any destructive or restructuring operation on a repo, and branch-scoped changes thereafter. An experiment is never the only copy of its result.
Practically: Python (scikit-learn, NumPy/pandas, matplotlib), detection tooling (Suricata, Zeek, YARA), Go, Rust and zsh for CLI work, Docker for lab isolation, Lean 4 and LaTeX when the problem is formal or the output is a document.
- Email — afiqandico13@gmail.com
- GitHub — @afiqandico
- LinkedIn — in/pangimpian
- X — @itsliszt
- ORCID — 0009-0002-1750-9551
Open to research collaboration on adversarial ML for network detection, detection-engineering work, and reproducibility tooling. For anything involving testing systems, I will ask for the scope in writing first.


