SDC4Society Tokyo — SDC Research at FHNW & swissanon

A short tour of our statistical-disclosure-control work: an R toolbox for anonymization, synthetic data and risk measurement; applied projects on longitudinal, health and mobility microdata; and new research on AI on both sides of the table — an AI that helps the anonymizer, and an agentic AI that re-identifies people from “anonymised” data. Thank you for hosting the conversation.

Slides (PDF) swissanon.com

The R toolbox

sdcMicro, simPop and riskutility are on CRAN; synvey is on GitHub (paper under review, CRAN release to follow).

sdcMicro (CRAN · GitHub) — the anonymization workhorse: risk measurement, recoding, suppression, PRAM, microaggregation, swapping; GUI sdcApp. On CRAN since 2007.
simPop (CRAN) — simulation of synthetic populations for SDC (design-aware, population size).
synvey (GitHub) — design-aware synthetic data generation at sample size; declares survey design (weights, strata, households) before synthesis. Paper under review.
riskutility (CRAN · GitHub) — ~60 disclosure-risk and utility measures under one interface, with a risk–utility map, for anonymized and synthetic data. Includes RAPID (Risk of Attribute Prediction-Induced Disclosure). RAPID method paper under review.

Projects

Currently running:

SNF Bridge (SNSF Bridge Discovery, 2023–2027) — privacy-preserving methods for event, longitudinal and trajectory data; risk assessment beyond k-anonymity.
Nestlé — anonymizing a small but complex health study (many sensitive variables, longitudinal follow-up) against identity and attribute disclosure.
SBB SynPop — household-aware anonymization and formal risk assessment for a nine-million-person, geo-referenced synthetic population.

Selected earlier work: World Bank, OECD, IHSN, Eurostat, Swisscom, Helsana.

AI in statistical disclosure control

AI as assistant SoftwareX · under review

An LLM proposes anonymization strategies inside sdcMicro, scores the information loss, and refines — a propose–evaluate–refine loop. Only variable-level metadata reaches the model, never records; it is provider-agnostic and runs fully on-premises with open-weight models. The practitioner signs off the generated R code.

AI as adversary PSD 2026 · to appear

A team of AI agents re-identifies people from “anonymised” mobility traces — cheaply, at scale, unsupervised. On a consenting-participant study it named the large majority of those who were publicly findable, for a few dollars and minutes per target. The lesson: de-facto anonymity is a moving line once the attacker’s effort collapses, and has to be re-assessed rather than assumed. To appear in Privacy in Statistical Databases (PSD) 2026, Springer LNCS.

Built to warn, under strict constraints: consent-only sample, simulated traces, public sources only, and non-release of both the dataset and the agent workflow.

Let’s work together

Idea 1 — a funded research & teaching visit

A stay of a few weeks up to ~2 months: an intensive lecture/seminar series on SDC, synthetic data and anonymization with R (sdcMicro, simPop/synvey, riskutility/RAPID), and joint research time to draft a paper and a grant proposal. The natural vehicle is a host-applied JSPS Invitational Fellowship (short-term 14–60 days or long-term 2–10 months): JSPS — not the host’s budget — covers the visitor’s travel, a living allowance, insurance and a research allowance. Formal teaching (with an honorarium) can be added via a host-university visiting professorship. Non-teaching time combined with family travel in Japan.

Idea 2 — a joint research project

A seed topic: SDC under agentic-AI adversaries — formal guarantees meet empirical red-teaming for mobility & population microdata. Two routes fit Switzerland + Japan, at different sizes:

SNSF–JSPS Strategic Japanese–Swiss Programme — bilateral (≤ CHF 300k / ≤ JPY 30M, 24–36 months), right-sized for two labs. The 2025 call has closed; we watch for the next.
Horizon Europe, Pillar II — Japan fully associable since January 2026. A spot-on topic (HORIZON-CL4-2026-04-DATA-06, synthetic data & compliant data use) just closed in April 2026; the nearest open one, SecureAI (HORIZON-CL3-2026-02-CS-ECCC-02, security & privacy of AI models), fits but closes 15 September 2026 — too soon to enter as a fresh consortium. Horizon also needs more partners (3+ countries, large consortia), so the realistic play is to build the consortium early and aim at the 2027 tranche.

Courses & services

swissanon offers open and in-house courses on anonymization and synthetic data, and works with organisations on defensible data sharing — “risk is measured, not asserted”. See swissanon.com.