Marina Gardella
Postdoctoral fellow at Centre Borelli - ENS Paris-Saclay
I am a researcher in image and document forensics, focusing on statistical methods to assess authenticity and detect anomalies in digital images. My work lies at the intersection of signal processing, mathematical modeling, and computer vision. My research is driven by the goal of bridging theory and real-world applications.
I am a postdoctoral fellow at Centre Borelli, in École Normale Supérieure Paris-Saclay. I got my PhD at the École Normale Supérieure Paris-Saclay in December 2023. I wrote my dissertation under the supervision of Dr. Miguel Colom, Pr. Jean-Michel Morel, and Pr. Pablo Muse. You can access a copy of my dissertation here. I earned my Master's degree at the Universidad de la República in Montevideo, Uruguay, in July 2019. I wrote my dissertation under the supervision of Pr. Elvio Accinelli. You can access a copy of my dissertation here. I got my B.S. in Mathematics in 2017, under the supervision of Dr. Jorge Iglesias, at the School of Sciences of the Universidad de la República, in Montevideo, Uruguay. I also earned my B.S. in Economics in 2016 from the School of Economics of the same University.
My research focuses on statistical approaches to image and document forensics. I study how traces left by image formation processes can be modeled and exploited to assess authenticity and detect anomalies. This perspective provides a unified framework to address problems ranging from manipulation detection to source attribution and AI-generated content.
I adopt a model-based approach in which an observed image is seen as the result of a sequence of transformations, including acquisition, optical effects, and processing steps. These processes leave measurable traces, such as sensor noise, optical artifacts, or compression patterns.
The goal is to characterize these traces through their statistical properties and to build models describing their expected variability under nominal conditions. Detection is then formulated as the identification of observations that significantly deviate from these models.
Manipulation detection. Detection of local inconsistencies in images and scanned documents.
Source identification. Analysis of intrinsic device signatures such as PRNU and DSNU to attribute images to their acquisition source.
AI-generated images. Study of statistical differences between synthetic images and physically acquired ones, including correlation structures and artifacts.
A contrario detection. Statistical frameworks for anomaly detection with explicit control of false alarms and interpretable decision criteria.