Davide Andreoletti

Sandra Mitrović, Davide Andreoletti, Omran Ayoub

ChatGPT has the ability to generate grammatically flawless and seemingly-human replies to different types of questions from various domains. The number of its users and of its applications is growing at an unprecedented rate. Unfortunately, use and abuse come hand in hand. In this paper, we study whether a machine learning model can be effectively trained to accurately distinguish between original human and seemingly human (that is, ChatGPT-generated) text, especially when this text is short. Furthermore, we employ an explainable artificial intelligence framework to gain insight into the reasoning behind the model trained to differentiate between ChatGPT-generated and human-generated text. The goal is to analyze model's decisions and determine if any specific patterns or characteristics can be identified. Our study focuses on short online reviews, conducting two experiments comparing human-generated and ChatGPT-generated text. The first experiment involves ChatGPT text generated from custom queries, while the second experiment involves text generated by rephrasing original human-generated reviews. We fine-tune a Transformer-based model and use it to make predictions, which are then explained using SHAP. We compare our model with a perplexity score-based approach and find that disambiguation between human and ChatGPT-generated reviews is more challenging for the ML model when using rephrased text. However, our proposed approach still achieves an accuracy of 79%. Using explainability, we observe that ChatGPT's writing is polite, without specific details, using fancy and atypical vocabulary, impersonal, and typically it does not express feelings.

Omran Ayoub, Davide Andreoletti, Aleksandra Knapińska et al.

Adapting to concept drift is a challenging task in machine learning, which is usually tackled using incremental learning techniques that periodically re-fit a learning model leveraging newly available data. A primary limitation of these techniques is their reliance on substantial amounts of data for retraining. The necessity of acquiring fresh data introduces temporal delays prior to retraining, potentially rendering the models inaccurate if a sudden concept drift occurs in-between two consecutive retrainings. In communication networks, such issue emerges when performing traffic forecasting following a~failure event: post-failure re-routing may induce a drastic shift in distribution and pattern of traffic data, thus requiring a timely model adaptation. In this work, we address this challenge for the problem of traffic forecasting and propose an approach that exploits adaptive learning algorithms, namely, liquid neural networks, which are capable of self-adaptation to abrupt changes in data patterns without requiring any retraining. Through extensive simulations of failure scenarios, we compare the predictive performance of our proposed approach to that of a reference method based on incremental learning. Experimental results show that our proposed approach outperforms incremental learning-based methods in situations where the shifts in traffic patterns are drastic.

Fatima Ezzeddine, Mirna Saad, Omran Ayoub et al.

Anomaly detection (AD), also referred to as outlier detection, is a statistical process aimed at identifying observations within a dataset that significantly deviate from the expected pattern of the majority of the data. Such a process finds wide application in various fields, such as finance and healthcare. While the primary objective of AD is to yield high detection accuracy, the requirements of explainability and privacy are also paramount. The first ensures the transparency of the AD process, while the second guarantees that no sensitive information is leaked to untrusted parties. In this work, we exploit the trade-off of applying Explainable AI (XAI) through SHapley Additive exPlanations (SHAP) and differential privacy (DP). We perform AD with different models and on various datasets, and we thoroughly evaluate the cost of privacy in terms of decreased accuracy and explainability. Our results show that the enforcement of privacy through DP has a significant impact on detection accuracy and explainability, which depends on both the dataset and the considered AD model. We further show that the visual interpretation of explanations is also influenced by the choice of the AD algorithm.

Davide Andreoletti, Aris Marcolongo, Natasa Sarafijanovic Djukic et al.

Accurately predicting potato sprouting before the emergence of any visual signs is critical for effective storage management, as sprouting degrades both the commercial and nutritional value of tubers. Effective forecasting allows for the precise application of anti-sprouting chemicals (ASCs), minimizing waste and reducing costs. This need has become even more pressing following the ban on Isopropyl N-(3-chlorophenyl) carbamate (CIPC) or Chlorpropham due to health and environmental concerns, which has led to the adoption of significantly more expensive alternative ASCs. Existing approaches primarily rely on visual identification, which only detects sprouting after morphological changes have occurred, limiting their effectiveness for proactive management. A reliable early prediction method is therefore essential to enable timely intervention and improve the efficiency of post-harvest storage strategies, where early refers to detecting sprouting before any visible signs appear. In this work, we address the problem of early prediction of potato sprouting. To this end, we propose a novel machine learning (ML)-based approach that enables early prediction of potato sprouting using electrophysiological signals recorded from tubers using proprietary sensors. Our approach preprocesses the recorded signals, extracts relevant features from the wavelet domain, and trains supervised ML models for early sprouting detection. Additionally, we incorporate uncertainty quantification techniques to enhance predictions. Experimental results demonstrate promising performance in the early detection of potato sprouting by accurately predicting the exact day of sprouting for a subset of potatoes and while showing acceptable average error across all potatoes. Despite promising results, further refinements are necessary to minimize prediction errors, particularly in reducing the maximum observed deviations.

Davide Andreoletti, Luca Luceri, Tiziano Leidi et al.

The importance of emotions in service and in product design is well known. However, it is still not very well understood how users' emotions can be incorporated in a product or service lifecycle. We argue that this gap is due to a lack of a methodological framework for an effective investigation of the emotional response of persons when using products and services. Indeed, the emotional response of users is generally investigated by means of methods (e.g., surveys) that are not effective for this purpose. In our view, Virtual Reality (VR) technologies represent the perfect medium to evoke and recognize users' emotional response, as well as to prototype products and services (and, for the latter, even deliver them). In this paper, we first provide our definition of emotion-driven services, and then we propose a novel methodological framework, referred to as the Virtual-Reality-Based Emotion-Elicitation-and-Recognition loop (VEE-loop), that can be exploited to realize it. Specifically, the VEE-loop consists in a continuous monitoring of users' emotions, which are then provided to service designers as an implicit users' feedback. This information is used to dynamically change the content of the VR environment, until the desired affective state is solicited. Finally, we discuss issues and opportunities of this VEE-loop, and we also present potential applications of the VEE-loop in research and in various application areas.

Davide Andreoletti, Omran Ayoub, Silvia Giordano et al.

The outbreak of coronavirus disease 2019 (covid-19) is imposing a severe worldwide lock-down. Contact tracing based on smartphones' applications (apps) has emerged as a possible solution to trace contagions and enforce a more sustainable selective quarantine. However, a massive adoption of these apps is required to reach the critical mass needed for effective contact tracing. As an alternative, geo-location technologies in next generation networks (e.g., 5G) can enable Mobile Operators (MOs) to perform passive tracing of users' mobility and contacts with a promised accuracy of down to one meter. To effectively detect contagions, the identities of positive individuals, which are known only by a Governmental Authority (GA), are also required. Note that, besides being extremely sensitive, these data might also be critical from a business perspective. Hence, MOs and the GA need to exchange and process users' geo-locations and infection status data in a privacy-preserving manner. In this work, we propose a privacy-preserving protocol that enables multiple MOs and the GA to share and process users' data to make only the final users discover the number of their contacts with positive individuals. The protocol is based on existing privacy-enhancing strategies that guarantee that users' mobility and infection status are only known to their MOs and to the GA, respectively. From extensive simulations, we observe that the cost to guarantee total privacy (evaluated in terms of data overhead introduced by the protocol) is acceptable, and can also be significantly reduced if we accept a negligible compromise in users' privacy.