Abstract
In the last ten years, Machine Learning (ML) and Artificial Intelligence (AI) have been experiencing a new “Renaissance”. Machines made extraordinary leaps in solving highly complex, cognitive tasks with superhuman performance. Image synthesis and generative modeling had an unprecedented impact, unleashing the possibility of realistically manipulating faces videos, known as Deepfakes or correctly matching a face among millions. At the same time, despite the abundant prior work on adversarial attacks and defenses, very little effort has been devoted to cease the dualism between perturbations as evil entities versus benign defense mechanisms. This project challenges this view and advances the hypothesis that adversarial machine learning may improve the state-of-the-art in privacy protection and in the preservation of human dignity against severe defamations such as those generated by Deepfakes or malicious face analysis tools. Our project twists the role of adversarial perturbations into benign tools. It will work in synergy along two directions: (i) Making Deepfake Detectors (DFD) robust against adversarial attacks to avoid evasion of the detection; (ii) Seeking principled methods for designing, crafting, and evaluating robust yet benign perturbations that disrupt Deepfake creation attempts or malicious face analytics tools. The project mission is to study, analyze and deliver perturbations that protect personal data from face manipulation, meanwhile reinforcing brittle Deepfake Detection systems. By defining the perturbations in a semantic space, rather than in the image space, perturbations can be made robust enough to be beneficial for various applications: to secure user content from AI-based cyberattacks or massive scraping using face recognition or to protect face videos against hyper-realistic face manipulations.
Dettagli del progetto
Responsabile scientifico: Giuseppe Lisanti
Strutture Unibo coinvolte:
Dipartimento di Informatica - Scienza e Ingegneria
Coordinatore:
"Sapienza" Universita' Di Roma(Italy)
Contributo totale Unibo: Euro (EUR) 91.800,00
Durata del progetto in mesi: 29
Data di inizio
28/09/2023
Data di fine:
28/02/2026
