My main research area consists in the development of Machine and Deep Learning solutions characterized by a "privacy-preserving" approach. The main tool used for this goal is Homomorphic Encryption (HE), a novel and complex family of encryption schemes which enables the processing of encrypted data (i.e., without the need to decrypt data before the processing happens).

Among the others, I focus in particular on the application of HE-DL to time-series analysis, forecasting, etc.

Publications

• EVAD: encrypted vibrational anomaly detection with homomorphic encryption
  Date: 2024-03-01
  Author(s): A. Falcetta and M. Roveri
  Venue: Neural Computing and Applications, Springer

  This work introduces Encrypted Vibrational Anomaly Detection (EVAD), a novel approach to privacy-preserving anomaly detection in industrial vibrational data using machine learning. By integrating one-class support vector machines with homomorphic encryption, EVAD enables the detection of equipment irregularities without compromising data privacy. This method is significant for its application in predicting equipment failures and ensuring the reliability of industrial operations, making it the first known privacy-secured solution in the literature for such purposes, proven effective on public datasets.

• TIFeD: a Tiny Integer-based Federated learning algorithm with Direct feedback alignment
  Date: 2023-10-25
  Author(s): L. Colombo, A. Falcetta, and M. Roveri
  Venue: 2023 International Conference on AI-ML Systems, Bangalore, India, 2023

  This paper presents TIFeD, a Tiny Integer-based Federated Learning algorithm that enables machine learning model training on devices with limited resources. TIFeD uses integer arithmetic and Direct Feedback Alignment for efficiency on such devices, diverging from traditional methods that rely on external cloud services for training. It facilitates both full-network and single-layer training modalities, allowing for flexible model training directly on-device. The approach is practical for tiny machine learning applications, with experimental results confirming its effectiveness. TIFeD is made available to the scientific community through a public repository.

• To Personalize or Not To Personalize? Soft Personalization and the Ethics of ML for Health
  Date: 2023-10-09
  Author(s): A. Falcetta, M. Pavan, S. Canali, V. Schiaffonati, and M. Roveri
  Venue: 2023 IEEE 10th International Conference on Data Science and Advanced Analytics (DSAA), Thessaloniki, Greece

  This paper introduces an approach to personalize Machine Learning (ML) models in health and medicine, focusing on ethical and epistemic considerations such as privacy, reliability, and trade-offs. The authors propose 'soft personalization', an ethically-informed framework that aims to limit personalization while respecting values like representativity, quality, non-maleficence, beneficence, and privacy. This interdisciplinary method suggests developing various models to choose from based on quality and safety, exemplified through a case study on glucose monitoring with privacy-preserving ML. The framework underscores the inevitability of trade-offs among values in personalized health technologies, highlighting the need for prioritizing certain values over others.

• GreenTea: Time-series Exploration as-a-Service for environmental science
  Date: 2023-06-05
  Author(s): F. Puoti, A. Falcetta, M. Roveri, D. Riva, and D. Chiggiato
  Venue: 2023 IEEE Conference on Artificial Intelligence (CAI), Santa Clara, CA, USA, 2023

  This paper introduces GreenTea, an integrated Cloud-streaming platform for time-series exploration and forecasting as-a-service for environmental data. Following the as-a-Service approach, GreenTea provides (possibly multiple concurrent) users expert in environmental science and pollution research (but not expert in the field of AI) with useful insights into several integrated environmental data sources, allowing them to leverage the exploration and forecasting/prediction abilities of machine and deep Learning algorithms. Through a Web App (or API), the users upload their data as well as Custom Code if needed receive back the results of the data exploration and forecasting provided by the platform. Effectiveness and efficiency of GreenTea have been successfully applied to environmental data of the Regional Environmental Protection Agency of Lombardy in Italy.

• CNAS: Constrained Neural Architecture Search
  Date: 2022-10-12
  Author(s): M. Gambella, A. Falcetta, and M. Roveri
  Venue: International Conference on Systems, Man, and Cybernetics (IEEE SMC), Prague, 2022

  This study introduces a neural architecture search (NAS) solution called Constrained NAS (CNAS) that is able to take into account constraints on the search for the neural architecture. CNAS is able to consider both functional constraints, such as the type of operations that can be performed in the neural network, and technological constraints, such as computational and memory demands. CNAS has been applied to Tiny Machine Learning and Privacy-Preserving Deep Learning with Homomorphic Encryption. This is the first time this solution has been proposed in the literature.

• T4C: A Framework for Time-Series Clustering-as-a-Service
  Date: 2022-09-19
  Author(s): A. Falcetta and M. Roveri
  Venue: 2022 CPS Summer School PhD Workshop, (CPSWS 2022), Pula, 2022

  This paper presents T4C, an open-source Python-based framework for time-series clustering-as-a-service. T4C integrates several time-series clustering models and techniques, and is able to generate websites for users to explore the results of the clustering on their uploaded time-series.

• TIMEX: an Automatic Framework for Time-Series Forecasting-as-a-Service
  Date: 2022-08-15
  Author(s): A. Falcetta and M. Roveri
  Venue: The Sixth International Workshop on Automation in Machine Learning, workshop of the 28th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (KDD), Washington D.C., 2022

  This paper introduces TIMEX, an automatic 'Time eXploration' framework for time-series forecasting-as-a-service. Unlike existing tools and libraries, TIMEX provides an automatic end-to-end forecasting pipeline that includes data ingestion, preprocessing, description, prediction, and service delivery. The paper also presents algorithmic advances in the automatic characterization of time-series relationships, automatic definition of the best training window, and automatic selection of feature transformation. TIMEX has been applied to the forecasting of the COVID-19 pandemic in Italy and to the prediction of two benchmark time-series; results are compared with a state-of-the-art solution.

• Privacy-Preserving Deep Learning With Homomorphic Encryption: An Introduction
  Date: 2022-08-01
  Author(s): A. Falcetta and M. Roveri
  Venue: IEEE Computational Intelligence Magazine

  This paper introduces privacy-preserving deep learning with homomorphic encryption (HE) and presents a methodology for designing privacy-preserving convolutional neural networks (CNNs). The methodology was applied to a privacy-preserving version of the LeNet-1 CNN and was successful on two image classification benchmarks. The paper also discusses the challenges and software resources available for privacy-preserving deep learning with HE.

• Privacy-preserving Time Series Prediction With Temporal Convolutional Neural Networks
  Date: 2022-07-19
  Author(s): A. Falcetta and M. Roveri
  Venue: The International Joint Conference on Neural Networks (IJCNN), Padova, 2022

  This paper presents a general privacy-preserving solution for time series prediction using homomorphic encryption (HE) and temporal convolutional neural networks (PINPOINT). The paper also introduces a cloud-based privacy-preserving system for forecasting-as-a-service based on the proposed PINPOINT models. Experimental results show the effectiveness of the proposed solution. This is the first time this solution has been proposed in the literature.

• A First Step Towards a Streaming Linked Data Life-Cycle
  Date: 2020-11-01
  Author(s): R. Tommasini, M. Ragab, A. Falcetta, E.D. Valle, S. Sakr
  Venue: International Semantic Web Conference (ISWC Satellites), Athens, 2020

  This paper explores the problem of handling Streaming Linked Data (SLD) and how to reuse existing resources to produce and consume streams on the Web. It covers the challenges of applying FAIR principles when publishing data streams and contextualizes the usage of prominent Semantic Web resources. The paper also provides three representative examples of real-world Web streams and open-sources the code used.

• A Privacy-Preserving Distributed Architecture for Deep-Learning-as-a-Service
  Date: 2020-07-19
  Author(s): S. Disabato, A. Falcetta, A. Mongelluzzo and M. Roveri
  Venue: The International Joint Conference on Neural Networks (IJCNN), Glascow, Scotland, 2020

  This paper introduces a distributed architecture for deep-learning-as-a-service that preserves user sensitive data while providing Cloud-based machine and deep learning services. It is tailored for Convolutional Neural Networks (CNNs) and utilizes Homomorphic Encryption. Experiments show that the proposed architecture is effective.

Other research activities

• Abstract at workshop: Privacy-preserving deep learning with homomorphic encryption
  Date: 2022-11-29
  Author(s): M. Roveri, A. Falcetta
  Venue: The 11th Italian Workshop on Machine Learning and Data Mining, workshop of the 21th International Conference of the Italian Association for Artificial Intelligence (AIxIA 2022) Udine, 2022

  The study proposes a methodology for designing privacy-preserving deep neural networks using homomorphic encryption. This method has been applied to privacy-preserving image classification and time series prediction.

• Tutorial: Privacy-preserving machine and deep learning with homomorphic encryption: an introduction
  Date: 2022-07-19
  Author(s): M. Roveri, A. Falcetta
  Venue: The International Joint Conference on Neural Networks (IJCNN), Padova, 2022

  A tutorial which presents an introduction to both Homomorphic Encryption and to its application on machine and deep learning algorithms.

Teaching activities

Undergraduate

• Informatica e Elementi di Informatica Medica
  Years: 2022, 2023
  Teaching assistant: basics of programming with the C language (50% of the module).

PhD

• Time-series exploration with Machine and Deep Learning: from theory to practice
  Years: 2021
  Teaching assistant: basics of time-series forecasting with Python.

Master Thesis

I am co-relator of 10+ Master thesis (Computer Science and Engineering).

Reviewer activity

• Neural Computing and Applications, Years: 2023
• IEEE Transactions on Artificial Intelligence, Years: 2023, 2022, 2021
• Expert Systems With Applications, Elsevier, Years: 2023, 2022, 2021
• Neural Networks, Elsevier, Years: 2022
• International Conference on Systems, Man, Cybernetics (IEEE SMC), Years: 2022
• International Conference on AI-ML Systems (AI-ML SYSTEMS), Years: 2023, 2022