Dr. Alessio Benavoli

Title

Securing Machine Learning from Threats Exploiting Unused Model Parameters

Summary

Applications are increasingly leveraging Deep Neural Networks (DNNs) to provide state-of-the-art per-
formance on a wide range of real-life problems. With the success Large Language Models (LLMs), new
generations of larger models are expected to become central to many systems that are integrated into hu-
man spaces and that permeate everyday life.
The virtues of over-parameterization have been established from both the statistical and computational
perspectives. However, this overparameterization leaves a large number of parameters that are not essen-
tial to the learned model and whose state is not important (they are "don"t cares"). In both software and hardware systems, undefined behavior and don"t-care states have been shown to be potential sources of vulnerabilities. In the same vein, we believe that unused parameters (UPs) represent undefined behavior and don"t-care states with respect to the primary model that can be exploited by attackers. Because of the substantially different nature of machine learning models, this undefined behavior manifests differently: as
extra capacity used to store information covertly, or as extra functionality to support unintended behavior when triggered. The goal of this project is to understand the security and privacy implications of the unused param-
eters present in machine learning models. The research explores fundamental questions and
algorithmic innovations for a class of machine learning exploits that targets the available spare capacity in
overparameterized models. We explore the problem both from the theoretical and empirical perspectives.
We also aim to explore these ideas and threat models in the context of other learning models including
federated learning, and generative models. Finally, we also aim to study potential mitigation, both
those relying on detection as well as elimination of malicious embedded data or functionality. The attacks
and optimizations, theoretical models and their analyses, generalizations and end-to-end attacks and proposed mitigations represent the intellectual merit of the project.

Funding Agency

Research Ireland

Date From

01/01/2026

Date To

31/12/2028

Title

Data analytics for the development of smart infusion pumps

Summary

Funding Agency

BD & SFI-Confirm

Date From

01/06/2021

Date To

01/06/2023

Title

Re-Make

Summary

The Re-Make project, coordinated by Politecnico di Milano, aims to implement, monitor, and characterize a bespoke sustainable high-performance repair and Additive Manufacturing (AM) technology based on solid state deposition. The project seeks to achieve functional products through a multi-faceted approach that combines multi-scale computational modelling with advanced experimental, analytical techniques, and data-driven approaches. Extensive background research has led to the definition 11 Doctoral candidates that will accomplish Re-Make"s plan through a list of realistically measurable and verifiable Research and Training Objectives.

Here at Trinity College, we developed the DC2 program, for the development of a new experimental set-up for impact high-speed visualizations with a high-speed camera. The findings of DC2 will provide with a deeper understanding of the impact process and the key particle/substrate interactions for a wide range of materials, and will allow to resolve in real-time the plastic and elastic deformation process during Cold Spray (CS).

Funding Agency

European Union Horizon Europe research and innovation programme

Date From

01/10/2023

Date To

30/09/2027

Title

Humans-in-the-Loop towards a more effective AI in manufacturing

Summary

The adoption of Artificial Intelligence (AI) for Process Optimisation (PO) in manufacturing can significantly contribute to make the Irish manufacturing sector more competitive and sustainable. However, its adoption has been slower than expected. Looking at the way AI is being used in PO, the issue is the need of tailoring AI-solutions for every machine tool and context (used material, shape of the manufactured object). With this approach, the industry will need tens of thousands of unique AI- models, which is clearly infeasible.

Instead of developing an AI-solution for each machine tool and context, HLOOP proposes to build a single AI-model exploiting the synergy between AI and human-workers. Assessing the quality of the manufacturing process is more effectively done by humans than AI, while AI is better at dealing with high-dimensional decision problems. We plan to use the feedback of machine operators to train on-the-fly (that is without using any pre-collected dataset) and in real-time an AI-model that can predict when the manufacturing process is good/bad, and then optimise it.

https://hloop.scss.tcd.ie/

Funding Agency

SFI

Date From

01/01/2023

Date To

30/06/2024