Insufficiency-Driven DNN Error Detection in the Context of SOTIF on Traffic Sign Recognition Use Case

Hacker, Lukas; Seewig, Jörg · 2023 · OpenAlex-citations

DOI: 10.1109/ojits.2023.3236531

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of ensuring safety in automated driving systems that rely on Deep Neural Networks (DNNs), specifically within the framework of Safety of the Intended Functionality (SOTIF) defined by ISO 21448. The authors identify that DNNs suffer from specific insufficiencies—such as lack of robustness, interpretability, and uncertainty representation—that traditional functional safety standards like ISO 26262 do not adequately cover. To mitigate risks from these insufficiencies, the study proposes a general, insufficiency-driven error detection approach capable of monitoring DNNs during runtime. The method aims to detect errors arising from in-distribution, out-of-distribution (OOD), and adversarial inputs by combining diverse monitoring techniques rather than relying on single-purpose methods. The methodology involves creating an error root cause model that links DNN-specific functional insufficiencies to triggering conditions and potential hazardous behaviors. Based on this model, the authors derive five general monitor categories: OOD Monitor, Saliency Monitor, Plausibility Monitor, Adversarial Monitor, and Uncertainty Monitor. Each category generates a specific error score reflecting the likelihood of an error related to a particular insufficiency. These scores are aggregated using a meta-model optimized via stack learning to estimate the overall probability of a triggering condition. The approach is implemented and evaluated on a traffic sign recognition (TSR) use case using self-created 3D driving scenarios simulated in MATLAB RoadRunner and Simulink. The experimental setup includes a DNN-based TSR function and the proposed insufficiency-driven detector, tested against various input types including standard, OOD, and adversarial data. The results demonstrate that the proposed approach effectively handles all tested error types, successfully detecting errors caused by in-distribution, out-of-distribution, and adversarial inputs. The study shows a performance benefit of this combined method compared to a baseline DNN without monitoring and against state-of-the-art individual DNN monitoring methods. By addressing multiple safety-related insufficiencies simultaneously, the approach provides a more comprehensive safety strategy than methods focusing on individual error root causes. The significance of this work lies in its integration of established SOTIF concepts, such as the cause and effect chain, into DNN runtime monitoring. It provides a structured framework for addressing the open-world problem in automated driving by systematically monitoring for DNN-specific failures. The findings suggest that combining diverse monitoring methods through a meta-model offers a robust solution for detecting DNN errors in safety-critical applications, thereby enhancing the reliability of AI-driven perception systems in automated vehicles. This contributes to the broader field by bridging the gap between theoretical DNN insufficiencies and practical safety monitoring strategies required for regulatory compliance and safe deployment.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• dms validation