Me in my homeplace, the Dolomites.
About Me
Hi!
I hold a PhD in Computer Science from the University of Padova, with an MSc in Theoretical Physics.
My PhD thesis work was supervised by Prof. Francesco Silvestri and co-supervised by Prof. Martin Aumüller. My research spans both theoretical and applied domains, with a core focus on Differential Privacy and Data Structures for high dimensional data.
Currently living in Oslo working as a AIoT researcher for Havguard AS, developing and deploying innovative IoT solutions and autonomous systems from seabed to space.
CV
Research Interests
- Differential Privacy
Is a mathematical definition that allows to quantify how much a randomized algorithm is private. For any two neighboring dataset \(D\sim D'\) where they differ on one user, a randomized algorithm \(\mathcal{M}\) is said to be \(\varepsilon\)-differentsially 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 (it is a max divergence).
The intuition is that by upper bounding the divergence between the realizion on \(D\) and \(D'\), any advesary cannot reliably distinguish between the two, making effectivetly the contribution of a single user private. This is usually achieve by masking the real statistics with additive noise, obtaining thus a privacy-utility trade-off.
- Randomized Algorithms & Data Structures
Do you know that if you have an algorithm that succeeds with probability strictly larger than $1/2$, it is sufficient to run that algorithm independently \(O(\log n)\) times and pick the most common answer to have a probability of success that is \(1-O(1/n)\)?
This is a standard result of concentration of measure in probability theory, and it is the foundational trick in many randomized algorithms like the Count Min-Sketch, which is essential in big data application as it allows to compute the frequency of a stream of data while keeping the space constant.
- Probability and Statistics
Do you know that if you draw \(N\) independent samples from a standard normal distribution, their maximum value is incredibly predictable? As \(N\) grows, the maximum of these \(N\) random variables sharply concentrates around \(\sqrt{2\log N}\). Surprisingly, the fluctuations around this value shrink to zero at a rate of \(O(1/\sqrt{\log N})\). This phenomenon is called super concentration as the variance vanishes as the number of random variable grows.
This is a standard result of extreme value theory, a tool I extensively used in a joint work with my supervisors Differentially Private High-Dimensional Approximate Range Counting, Revisited.
- Deep Learning
Do you know that training a machine learning model on private data can be problematic due to the phenomenon of memorization? Generative AI algorithms may regurgitate private data if no precautions are taken. One way to mitigate this effect is to perform a differentially private training by adding noise to the stocastic gradient descent
\[\theta_{t+1} = \theta_t - \eta {1\over L}\bigg(\sum_{i\in B}\text{clip}_C(\nabla_\theta \mathcal{L}(\theta_t, x_t)) + \mathcal{N}(0, \sigma^2C^2 \, 1)\bigg).\]A summary of the state of the art (up to 2024) can be found in my project The Privacy Analysis of the Differential Private Stochastic Gradient Descent.
There are well know libraries that provide good wrapper to TensorFlow and Pytorch for differentially private training. In of the is Opacus.
Previous Research Interests
- Complex Systems (MSc thesis in Economic Complexity)
- Mobility Data Science (PhD collaboration with Motion Analytica)
- Quantum Computing
- Statistical Physics
Academic & Professional Experience
- AIoT Researcher - Havguard AS - Oslo. Engineered and adapted predictive algorithms in C++ for resource-constrained environments, focusing on high-performance and system stability.
- Research Background. Prior to my PhD, I completed my studies in Theoretical Physics at the University of Padova, which provides the mathematical foundation for my current work in randomized algorithms.
- Visiting Research. In 2024, I was a visiting PhD student at the IT University of Copenhagen, working with Prof. Rasmus Pagh as part of the Providentia project and my co-supervisor Prof. Martin Aumüller.
- Industry Collaboration. I collaborate with Motion Analytica srl on mobility research. Our collaborative work was recently published in a peer-reviewed journal link.
Beyond the Lab
In my spare time, I prioritize staying active and creative:
- Hiking: I am a hiking guide assistant for The South Adventure. I recently completed the Annapurna Circuit trek in Nepal reaching 5400 meters above sea level. Hiked the highest mountain in North Africa, the Jbel Toubkal.
- Music: I play bass, guitar, and a bit of piano. I was previously part of a band and contributed to the album L’assenzio (Spotify).
- Composition: I enjoy writing and producing music; you can listen to some of my tracks on SoundCloud.
- Fitness: I love calisthenics.
Posts
subscribe via RSS