The last decade has seen a surging opportunity for developing autonomous systems that require making decisions without harming the health of people, infrastructures, environment, or causing financial losses. Unfortunately, taking decisions autonomously through the support of Machine Learning classifiers to address complex tasks, is far from trivial and not ready yet to deal with highly critical tasks. This paper advocates the need to focus on trust rather than traditional correctness, embracing software design patterns proven in use for decades, and refurbishing them considering current needs. This paper introduces Self-Checking Classifiers (SCCs), governed by an Acceptable Level of Risk (ALR), which transform uncertain predictions into rejections to bound misclassification in safety-critical contexts. We further show that redundant SCC pairs can reduce unnecessary rejections compared to single SCCs, improving availability under the same risk constraints. The proposal is formalized and then evaluated using various classifiers, uncertainty estimates and rejection strategies. The discussion, findings and framework presented in this paper will support the design, deployment, and benchmarking of trustworthy decision-making software, empowering future (autonomous) critical software and systems.

Back to the Future: Refurbishing Critical Software Engineering to Enable Trustworthy Classification / Zoppi T., Atif M., Khokhar F.A., Bondavalli A.. - ELETTRONICO. - (2026), pp. 359-368. (41st Annual ACM Symposium on Applied Computing, SAC 2026 Grand Hotel Palace, grc 2026) [10.1145/3748522.3779789].

Back to the Future: Refurbishing Critical Software Engineering to Enable Trustworthy Classification

Zoppi T.
;
Atif M.;Khokhar F. A.;Bondavalli A.
2026

Abstract

The last decade has seen a surging opportunity for developing autonomous systems that require making decisions without harming the health of people, infrastructures, environment, or causing financial losses. Unfortunately, taking decisions autonomously through the support of Machine Learning classifiers to address complex tasks, is far from trivial and not ready yet to deal with highly critical tasks. This paper advocates the need to focus on trust rather than traditional correctness, embracing software design patterns proven in use for decades, and refurbishing them considering current needs. This paper introduces Self-Checking Classifiers (SCCs), governed by an Acceptable Level of Risk (ALR), which transform uncertain predictions into rejections to bound misclassification in safety-critical contexts. We further show that redundant SCC pairs can reduce unnecessary rejections compared to single SCCs, improving availability under the same risk constraints. The proposal is formalized and then evaluated using various classifiers, uncertainty estimates and rejection strategies. The discussion, findings and framework presented in this paper will support the design, deployment, and benchmarking of trustworthy decision-making software, empowering future (autonomous) critical software and systems.
2026
Proceedings of the ACM Symposium on Applied Computing
41st Annual ACM Symposium on Applied Computing, SAC 2026
Grand Hotel Palace, grc
2026
Zoppi T.; Atif M.; Khokhar F.A.; Bondavalli A.
File in questo prodotto:
File Dimensione Formato  
3748522.3779789.pdf

accesso aperto

Tipologia: Pdf editoriale (Version of record)
Licenza: Open Access
Dimensione 561.34 kB
Formato Adobe PDF
561.34 kB Adobe PDF

I documenti in FLORE sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificatore per citare o creare un link a questa risorsa: https://hdl.handle.net/2158/1499895
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? ND
social impact