|BONETTINI NICOLÒ||Cycle: XXXIV |
Tutor: MONTI-GUARNIERI ANDREA VIRGILIO
Advisor: MARCON MARCO Major Research topic
:Signal processing for enhanced data-driven anomaly detectionAbstract:
My research project focuses on combining Signal Processing with data-driven techniques for anomaly detection.
The goal is to develop a new class of AI-based tools, that can be successfully used to preserve the dependability of digital data. To do so, we pursue a novel approach whereby AI-techniques are enriched with a model-based signal-processing viewpoint. In this way, the strengths of data-driven techniques are maintained while, at the same time, exploiting any available information stemming from the scenario at hand.In this vein, the need for huge amount of training data tipically required by data-driven techniques is relaxed by constraining the model and the training process so to orient the analysis towards task-relevant features.
A possible application of the aforementioned tools is Food Contaminants detection by Hyperspectral X-ray Imaging. The typical hyperspectral X-ray imaging pipeline is composed of a hyperspectral X-ray generator and a detector. X-rays issued by the generator go through food containers on a conveyor belt and are measured by a hyperspectral detector. In such a scenario, few relevant information (i.e. few X-ray energy bins) is hidden in large amount of high-dimension data. Being able of highlighting relevant information in pre-processing phase and using a hybrid approach is paramount to be effective in anomaly detection and classification problems.
Another possible application is in Multimedia Forensics (MMF) field. A tipical MMF signal is a picture or a video with plenty of non-relevant scene content and very subtle traces (e.g. camera sensor noise) we would be able to exploit. The risk of using pure data-driven machine learning approach is to learn unnecessary features for the considered task, leading to poor models and poor results. This risk is minimized by using a hybrid approach (i.e. both data and model driven) to focus on specific traces.