Some of the topics covered are the following: Edge-Vision and Efficient deep learning, Multimedia Forensics and Deepfake detection, Deep Learning for image and video analysis, Visual Knowledge acquisition: Activity Recognition & Object Detection, Computer Graphics and Point cloud representation, Monocular Depth Estimation, Energy aware deep learning models and Green AI (Beekeeping and Vertical farming), Adversarial Machine Learning.
Lightweight DL models for resource constraints device — Investigating both mathematically and practically, research areas related to the efficiency of DL models for computer vision tasks. More in detail, the objective is to theoretically analyze and investigate the behavior of fundamental components for neural network learning mechanisms, with a focus on specific layers and elements that characterize the learning procedure, such as self-attention, knowledge distillation, and optimizers. More in detail, computationally efficient solutions are developed in the fields of perception and security, i.e., studying efficient techniques in well-known tasks like monocular depth estimation, 3D mesh reconstruction, and deepfake detection. Additionally, the key elements of neural network efficiency is analyzed, such as inference time, energy consumption, and their trade-off with estimation performances.
Multimedia forensics and deepfake detection — Multimedia Forensics includes a set of scientific techniques recently proposed for the analysis of multimedia signals (audio, videos, images) in order to recover probative evidences from them; in particular, such technologies aim to reveal the history of digital contents, such as identifying the acquisition device that produced the data, validating the integrity of the contents and retrieving information from multimedia signals. With our research, we seek to study these models to create defense solutions against disinformation attacks based on diffusion models and generative techniques..
Action and Activity Recognition, Anticipation and Forecasting — Different works in literature afford the problem of Actions and Activities Recognition, Anticipation and Prediction in videos. The complexity of the problem requires the consideration of many aspect. First of all, the recognized action sequence has to be consistent with the final task of the whole activity. Furthermore, much attention needs to be given to the prediction of the correct action in those instances where specific sequences are under represent in the dataset not because of the likelihood of them to happen. Finally, several implementation problems, caused by the large dimension of the data used, need to be addressed. Our researched work focused on tackling those problems producing a novel network, the Anticipation and Forcasting Network.
Object Detection and Instance Segmentation — Object detection is the task of detecting instances of certain object classes (such as humans, buildings or cars) in digital images and videos. Well-researched sub-tasks include face detection and pedestrian detection. Instance segmentation is the task of grouping parts of the image that belongs to the same entity or class. In the field of research that combines Object Detection and Instance Segmentation, a new approach is proposed: from the classical machine learning algorithms, the research community moved to a neu ral network approach via the use of several new architecture. Our research focused on developing new architectures by improving performances, computation time, capacity and multi-tasking properties.
Edge and Fog Computing — Distributed algorithms are stutied, resource-sharing strategies, and scheduling policies that best fit this new kind of computing paradigm. We have expertise in the mathematical modeling of the problem, simulation tools and in modern technologies like Docker and Kubernetes.
Impact of adversarial and backdoor attacks on deep learning techniques — In the last couple of decades, machine learning and neural networks applications have quickly become the state of the art in every automated task. Moreover, the spreading of high-performance GPUs at an affordable price, along with the creation of frameworks that are always simpler to use, have made the implementation of neural network architectures accessible to everyone. Nevertheless these techniques have been found to be highly exposed to malicious approaches. Malicious approaches usually refer to an adversarial scenario, in which an attacker tries to exploit vulnerabilities of a system in order to gain advantages from it.
