Current students

Assessing the integrity of a multimedia object is matter of study of the multimedia forensics (MMF) community. This is a problem that has been historically tackled modelling some non-invertible traces left by editing operations, called forensic footprints (FFs), through classical and powerful signal processing techniques.

Nowadays, the MMF community has gradually switched towards more data-driven approaches. These methods do not incorporate FFs knowledge, still they outperform classic techniques on all types of contents and a variety of tasks. This has been a necessary step given the rate at which new tampering methods emerge in the scenario of modern Internet-based communication systems (e.g., social media, chat services, etc.).

The drawback for this last family of methods is that they often lack some desirable properties for a forensic detector. For instance, they lack interpretability, i.e., knowing what prompted a detector to a specific decision. Moreover, it is often unclear how to assess the confidence of the decision taken (e.g., if an image is detected as manipulated, how can we quantify how sure the detector is about this decision?).

In sensible areas like MMF, it is extremely useful to understand which are the forensic clues exploited by a detector and which is the uncertainty behind the detector response in case of either a positive (correct decision) or negative (incorrect decision) outcome. Furthermore, results interpretability can give more insight on the FFs that new, never seen tampering pipelines introduce in forged objects.

Given that interpretability of a detector response in the MMF context is not straightforward as many FFs are difficult to spot in the original input domain, the goal of my thesis is to bridge the gap between the high-performances offered by data-driven techniques and their flaws when dealing with current forensic needs. This is a multi-faceted objective that will be achieved in several ways, for instance: integrating FFs knowledge from classic approaches into data-driven pipelines; defining and implementing interpretable forensic detectors; employing uncertainty estimation techniques to assess the reliability of the analysis executed; finally, studying how these techniques can be applied to the problem of generalization when dealing with semantically similar tampering techniques, i.e., data with similar semantic content but processed with different pipelines.