--- title: "Federated Learning, Trusted Third-Party Custody & Differential Privacy" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Federated Learning, Trusted Third-Party Custody & Differential Privacy} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) ``` ## Overview `dicomtk.fl` enables multi-institutional imaging consortia to collaborate without centralising patient-level DICOM or clinical identifiers: 1. **Trusted Third-Party (TTP) & Polymorphic Pseudonymisation (`PEP`)**: `ttp_mainzelliste_client`, `pep_derive_site_token()`, and `pep_rekey_pseudonym()` issue unlinkable project- and site-specific pseudonyms. 2. **Privacy-Preserving Record Linkage (`PPRL`)**: `pprl_encode_bloom()`, `pprl_dice_similarity()`, and `pprl_link_cohorts()` match patients across hospitals using frequency-hardened Bloom filters. 3. **Differential Privacy (`DP`) Federated Statistics**: `dp_budget_accountant`, `dp_suppress_small_cells()`, `dp_federated_summary()`, and `dp_aggregate_sites()` release bounded $(\varepsilon, \delta)$-DP cohort summaries while enforcing $k$-anonymity. 4. **Governed FL Site Package Exports**: `fl_site_approval_gate` and `fl_export_site_package()` require cryptographic `site_admin` sign-off before emitting Flower, NVIDIA FLARE, or OpenFL bundles. ```{r example} library(dicomtk.fl) # 1. TTP + Polymorphic Pseudonymisation ttp <- ttp_mainzelliste_client$new(mode = "offline", ttp_secret = "demo-secret") sess <- ttp$create_session() root_id <- ttp$request_pseudonym( sess$session_id, list(vorname = "Alice", nachname = "Tan", geburtsdatum = "1985-04-12") ) tok_a <- pep_derive_site_token(root_id$pid, "PROJ_A", "SITE_1", "pep-key") tok_b <- pep_rekey_pseudonym(tok_a, "PROJ_B", "SITE_2", "pep-key") tok_a$pseudonym != tok_b$pseudonym # 2. DP Federated Summary acc <- dp_budget_accountant$new(max_epsilon = 3.0, max_delta = 1e-5) set.seed(1) s1 <- dp_federated_summary(rnorm(120, 62, 8), lower = 0, upper = 120, accountant = acc, site_id = "SITE_1") s2 <- dp_federated_summary(rnorm(150, 65, 8), lower = 0, upper = 120, accountant = acc, site_id = "SITE_2") dp_aggregate_sites(list(s1, s2)) ```