Seminari per la chiamata di tre RTDA
Con riferimento alla chiamata di due Ricercatori a Tempo Determinato di tipologia A per il Settore scientifico disciplinare INF/01, si comunica che in data 17 marzo 2023, su Google Meet si terranno i seguenti seminari:
- alle ore 10:00 il Dott. Indro Spinelli, vincitore del Bando RTDA PNRR n. 1-2023
Titolo: Trustworthy Graph Neural Networks
Abstract: To be trusted from a human perspective requires robustness from malign input manipulation, privacy protection of the ingested data, fair treatment for every represented group of individuals and transparency in the decision process. However, the "vanilla" formulation of graph neural networks (GNNs) and do not possess these characteristics. In this seminar, we focus on fairness and interpretability. The tendency of similar nodes to cluster on several real-world graphs (i.e., homophily) is the first cause of unfairness. We tackle the culprit by proposing a biased pruning of the graph connections to reduce the homophily of the sensitive traits. Then, addressing the drawbacks of the previous solution, we directly learn to rewire the graph's topology to make it fair. The explainability goal is to associate additional information to allow human experts to interpret and extract knowledge from the model. We develop a meta-learning framework for improving the level of explainability of a GNN at training time by steering the optimization process towards an "interpretable" local minima. Finally, we exploit the recent trend of training GNNs over a bag of sub-graphs. However, in our case, the bag is built with the double objective of improving the predictive performances of the model and providing an easy-to-interpret explanation. The holy grail of a trustworthy model requires the presence of all these characteristics at once without downgrading the model's utility. Therefore, it is safe to assume that an exciting period awaits us in this research field. - alle ore 10:30 il Dott. Dorjan Hitaj, vincitore del Bando RTDA PNRR n. 2-2023
Titolo: Machine Learning meets Cybersecurity (and vice versa)
Abstract: With the boom of machine learning in the past decade, its use to improve the performance of cybersecurity systems became obvious. Cybersecurity systems employing machine learning techniques can analyze potential threat patterns, reason about past events, and adapt to evolving information. More importantly, machine learning models can aid cybersecurity analysts by suggesting various mitigation strategies, resulting in a rapid and swift response towards attacks. But At the same time, a growing body of work has demonstrated that machine learning models are susceptible to attacks compromising both the security and privacy guarantees of these models. While there does exist a tremendous amount of work proposing defense mechanisms, the concern regarding the robustness and reliability of machine learning solutions, especially in critical cybersecurity systems, is still present. - alle ore 10:30 il Dott. Daniele Friolo, vincitore del Bando RTDA PNRR n. 3-2023
Titolo: Privacy in Distributed Ledger Technologies: a cryptographic challenge.
Abstract: Distributed Ledger Technology, or Blockchain, is largely used nowadays for different applications like cryptocurrencies, decentralized finance, supply chains, decentralized gaming. Modern blockchains like Ethereum or Algorand are based on an enhanced technology called Smart Contracts, which is a piece of code posted on the distributed ledger that establishes a publicly-verifiable contract between two or multiple parties. Smart contracts are often expressed with Turing-complete languages and are usually executed by a virtual machine mounted on top of the thousand nodes keeping the distributed ledger alive. Such nodes will agree on the functions’ output upon consensus. Today privacy in DLT is a major concern, and making thousands of untrusted nodes compute a smart contract function on private data in a scalable way is a huge challenge for the cryptographic community. The presented research covers different aspects of privacy in DLT, and shows how modern cryptographic tools like multi-party computation and zero-knowledge proofs can be used to obtain privacy guarantees on top of the already used DLT architectures like Ethereum.