Resilient Computing Lab
Possible subjects of Bachelor, Master and PhD Theses
The RCL group continuously supports Bachelor, Master and PhD students for their Theses. In general, vast majority of our arguments refer to the field of critical systems i.e., systems whose failure may lead to damage, injuries or economic losses. Some possible research challenges that can be investigated during the Thesis, and contextualixed in their application domains, are reported below.
Safety of machine learning solutions. Machine learning is undeniably an enabling technology in several domains, for example it is at the foundation of autonomous driving. However, unproper (unsafe) behaviour of solutions based on machine learning may lead to dangerous consequences. Many Theses initiatives at the MSc, BSc, and PhD level can be identified in this direction, especially using the autonomous driving as a reference domain, and that may span from the definition of mitigation strategies of possible unsafe behaviours to the comparison of solutions or to their representation and assessment through simulators.
Unsafe object detectors for safe autonomous driving. The purpose is to explore mechanisms to achieve safe autonomous driving in the presence of (unsafe) object detectors. Autonomous driving relies heavily on the output produced by object detectors, for example for trajectory planning. Despite the object detectors may misdetect objects, it is required that the driving task is safe, avoiding hazardous manouvers that may lead to accident. Multiple initiatives at the MSc, BSc, and PhD level can be identified in this direction, especially focused on understanding the impact that misdetection may have to the trajectory planning taks, and architectural or algorithmical solutions to mitigate the impact of misdetection.
Identification of intrusions and failures through monitoring. Monitoring data streams allows detecting attacks as well as possible failures of software or hardware components. In particular, anomaly detection is a machine-learning solution that allows identifying deviations from the expected behaviour; for example, several research works and commercial tools exploit anomaly detection to identify cyber-attacks. Several ideas for possible research can be proposed to interested Students, from the implementation and assessment of algorithms, to their proper tuning and combination with the aim to improve detection scores.
Design, implementation and validation of standard-compliant systems. Critical systems are typically crafted following prescriptions written in standards. The aim of these standards is to enforce on the system specified properties as safety (e.g., ISO 61508 for electronic equipment) or security (e.g., ISA/IEC 62443 for industrial automation and control systems). Since long, there is a constant strive to identify and exercise methodologies, techniques and tools that can provide greater and greater evidence of compliance to the desired properties and at lowered costs. The RCL groups is regularly involved in projects and is in close contact with companies where the design, implementation and evaluation of components is done following such standards, and trying to improve the existing state of the art on methodologies and techniques.
Modeling of Critical Systems. Modeling is an art, which consists of capturing the important components and relationships of a system while removing unimportant details. Models can be used both for designing a system, and for analyzing its dependability, security, performability properties. Possible topics of Bachelor/Master’s thesis include, but are not limited to: 1) Development/application of Model-Driven-Engineering approaches for modeling and analyzing critical systems; 2) Development/application of stochastic state-based modeling approaches for quantitative assessment of dependability, security, performability attributes of critical systems. Possible topics/technologies involved: UML profiling, Eclipse Modeling Framework, eCore models, languages/tools for Model-to-Model and Model-to-Text transformations (ATL, Viatra, EGL, ...), formalisms for stochastic modeling (SPN, SAN, …), supporting analysis tools (Möbius, …).
The list of topics is not exhaustive, and also new topics continuously emerge. The most effective way is to approach members of the RCL groups for updated information.
Some precise topics currently available are the following:
Object criticality to improve safety of trajectory planning. Object detection in autonomous driving consists in perceiving and locating instances of objects in multi-dimensional data, such as images or lidar scans. Very recently, multiple works are proposing to evaluate object detectors by measuring their ability to detect the objects that are most likely to interfere with the driving task. Detectors are then ranked according to their ability to detect the most relevant objects, rather than the highest number of objects. However there is little evidence so far that the relevance of predicted object may contribute to the safety and reliability improvement of the driving task. For this reason, an additional parameter, the predicted criticality of an object, has been proposed and means to compute it has been implemented. Predicted criticality has been used in conjunction with prediction confidence to facilitate the task of trajectory planning. However, a proper configuration of such parameters require additional studies. Starting from available results and implementations, the Student is requested to experiment with the above, to identify proper configurations that maximizes performances of trajectory planners.
On the effect of prediction errors in cyber-physical systems. It is well-known that machine learners may make wrong predictions. Cyber-physical systems encompassing machine learners may use those predictions to take erroneous actions. The research work aims to explore different applications where machine learning is exploited, identifying possible scenarios where the machine learner makes mistake, and the likely consequences at system level. Each scenarios, including its physical environment, is detailed to identify the conditions under which a system failure occurs, and, as a direct consequence, the operating condition of the machine learner under which the system is expected to behave properly. From these considerations, general rules to define target requirements of a machine learner operating in a determined physical environment are finally raised. To restrict the scope of the research and facilitate the student, the analysis will focus on the autonomous driving domain.
Modelli per l’analisi della sicurezza di sistemi critici. La tesi è incentrata sugli approcci di modellizzazione per l’analisi della sicurezza di sistemi critici. Lo studente sarà inizialmente introdotto ai concetti di base riguardanti l’affidabilità e la sicurezza dei sistemi informatici e agli approcci di modellizzazione per l’analisi quantitativa dei sistemi. In seguito saranno approfonditi alcuni approcci di modellizzazione caratterizzati dalla loro capacità di catturare esplicitamente il comportamento di un attaccante del sistema, quali gli Attack trees, gli Attack graphs, i privilege graphs, ed altri formalismi quali ADVISE (ADversary VIew Security Evaluation). Particolare attenzione sarà posta nell’andare a identificare e sistematizzare le caratteristiche di ciascun approccio, e nel verificare la presenza di eventuali tool a supporto.
Sviluppo, revisione ed integrazione di un framework di modellizzazione. Lo scopo principale della tesi è quello di sviluppare, aggiornare ed integrare i vari componenti del framework TAME (Template Modeling Environment - Link) già parzialmente sviluppati (Link1). Il focus del framework è la realizzazione di modelli template-based per l’analisi della performability di sistemi. Dopo una prima fase in cui verrà presa confidenza con il framework (Link2), lo studente dovrà quindi andare a correggere i problemi di implementazione precedentemente riscontrati al fine di integrare i vari componenti del TAME. Per fare ciò verrà utilizzato l’ambiente Eclipse, in particolare il pacchetto EMF (Eclipse Modeling Framework), insieme a varie features disponibili (Ecore, ATL, Epsilon).
Applicazione del framework di modellizzazione TAME. L’obiettivo della tesi è quello di applicare il framework sperimentale TAME (Template Modeling Environment) a casi di studio di sistemi dependable. TAME è un framework template-based per la modellizzazione di sistemi e l’analisi della performability. Lo studente dovrà quindi studiare e prendere confidenza con il framework ed applicarlo per modellizzare ed analizzare un caso di studio concreto. La modellizzazione verrà effettuata attraverso il formalismo SAN-T (Stochastic Activity Network Templates - Link), un’estensione delle SAN con l’aggiunta di elementi di variabilità.
Algoritmi Supervised/Unsupervised e loro Capacità di Rilevare Anomalie Sconosciute. La tesi si pone l'obiettivo di studiare algoritmi di Machine Learning che sono comunemente utilizzati per la rilevazione di Anomalie (classificazione binaria) allo scopo di misurare quantitativamente la loro capacità di rilevare anomalie sconosciute. Lo studente dovrà prima padroneggiare i concetti di base di un processo di apprendimento automatico (Machine Learning), per poi derivare una lista di algoritmi supervisionati (Supervised) e non supervisionati (Unsupervised) che vengono comunemente utilizzati come classificatori. Tali algoritmi dovranno poi essere eseguiti su diversi dataset (forniti dal docente) allo scopo di calcolare delle metriche quantitative che permettono di valutare le capacità di classificazione di anomalie sconosciute. La metodologia per l’esercuzione di questa campagna di esperimenti sarà derivata ed implementata (usando tool disponibili ed eventualmente con piccoli task di programmazione) dallo studente durante il lavoro di tesi, assieme ad una campagna di analisi dati atta a presentare e confrontare i risultati ottenuti.