Fabrizio Boninsegna

Research interests

Differential privacy

A mathematical definition that lets you quantify how private a randomized algorithm is. For any two neighbouring datasets \(D \sim D'\) differing in a single user, a randomized algorithm \(\mathcal{M}\) is \(\varepsilon\)-differentially private if

\[\text{Pr}[\mathcal{M}(D) = y] \leq e^\varepsilon\, \text{Pr}[\mathcal{M}(D') = y].\]

It is a measure of divergence between two probability distributions — a max divergence. The intuition: by upper bounding the divergence between the outputs on \(D\) and on \(D'\), no adversary can reliably distinguish the two, which makes a single user’s contribution effectively private. This is usually achieved by masking the true statistics with additive noise, which is where the privacy–utility trade-off comes from.

Randomized algorithms and data structures

Did you know that if you have an algorithm succeeding with probability strictly greater than \(1/2\), it is enough to run it independently \(O(\log n)\) times and take the majority answer to succeed with probability \(1 - O(1/n)\)?

That is a standard concentration-of-measure result, and it is the foundational trick behind many randomized algorithms — the Count-Min sketch among them, which computes frequencies over a data stream in constant space.

Probability and statistics

Did you know that if you draw \(N\) independent samples from a standard normal distribution, their maximum is remarkably predictable? As \(N\) grows, the maximum concentrates sharply around \(\sqrt{2 \log N}\), and the fluctuations around that value shrink at a rate of \(O(1/\sqrt{\log N})\). The phenomenon is called super-concentration: the variance vanishes as the number of random variables grows.

This is a standard result of extreme value theory, a tool I used extensively in joint work with my supervisors, Differentially Private High-Dimensional Approximate Range Counting, Revisited.

Deep learning

Did you know that training a machine learning model on private data is risky because of memorization? Generative models can regurgitate private data if no precautions are taken. One way to mitigate this is differentially private training, adding noise to stochastic gradient descent:

\[\theta_{t+1} = \theta_t - \eta \frac{1}{L}\bigg(\sum_{i \in B}\text{clip}_C\big(\nabla_\theta \mathcal{L}(\theta_t, x_t)\big) + \mathcal{N}(0, \sigma^2 C^2 \mathbf{I})\bigg).\]

A summary of the state of the art (up to 2024) is in my project report, The Privacy Analysis of the Differentially Private Stochastic Gradient Descent. Well-known libraries wrap TensorFlow and PyTorch for private training — Opacus is one of them.

Earlier interests

Complex systems (MSc thesis in economic complexity) · mobility data science (PhD collaboration with Motion Analytica) · quantum computing · statistical physics.

Experience

AIoT Researcher — Havguard AS, Oslo

Engineering and adapting predictive algorithms in C++ for resource-constrained environments, with a focus on high performance and system stability.

Visiting PhD student — IT University of Copenhagen, 2024

Worked with Rasmus Pagh as part of the Providentia project, together with my co-supervisor Martin Aumüller.

Industry collaboration — Motion Analytica

PhD research collaboration on mobility research with Motion Analytica. Our joint work was recently published in a peer-reviewed journal.

Background

Before the PhD I studied theoretical physics at the University of Padova, which is where the mathematical foundation for my work on randomized algorithms comes from.

Beyond the lab

• Hiking. I am a hiking guide assistant for The South Adventure. I recently completed the Annapurna Circuit in Nepal, reaching 5 400 m, and hiked Jbel Toubkal, the highest mountain in North Africa.
• Music. I play bass, guitar and a bit of piano. I was part of a band and contributed to the album L’assenzio.
• Composition. I write and produce music — some tracks are on SoundCloud.
• Fitness. I love calisthenics.