The IEEE Industry Connections 2025 report on structured synthetic data, co-authored by Matteo Giomi, examines the legal, technical, and governance dimensions of synthetic data and establishes that anonymity cannot be assumed from generation alone.
Perspectives on AI systems, privacy, and evaluation from the Modal Resonance team.
RSS FeedThe IEEE Industry Connections 2025 report on structured synthetic data, co-authored by Matteo Giomi, examines the legal, technical, and governance dimensions of synthetic data and establishes that anonymity cannot be assumed from generation alone.
Applying Anonymeter to over 150 de-identified datasets from the NIST Collaborative Research Cycle shows that no current synthesis algorithm achieves residual privacy risk below a few percentage points, and that algorithm choice can change measured risk by an order of magnitude.
Anonymeter is an open-source framework for measuring privacy risk in synthetic tabular data, published at PoPETs 2023 and evaluated positively by France's CNIL.
We benchmarked three open-source PII detectors (Microsoft Presidio, GLiNER, and OpenAI's Privacy Filter) on detection accuracy and throughput across two datasets and six entity types. OPF leads out-of-the-box on a GPU, Presidio wins on CPU and pattern-rich entities, GLiNER is the most flexible to configure.
A structured breakdown of where privacy risk lives across the four phases of the AI lifecycle, and what genuinely changes with generative AI versus traditional ML.