Guideline for AI in Safety-Critical Applications

Click on the image to enlarge

The mobility of the future becomes more and more autonomous. Artificial intelligence is a key technology in this context.

However, its utilization may lead to new threats, which need to be mitigated. Existing safety concepts are no longer suitable. The "Guideline for AI in safety-critical applications" shall contribute to a safe and autonomous mobility.

Guiding Principles

This Guidelines contains a list of principles to enable the safe development and deployment of Artificial Intelligence. The principles address AI specific challenges and insufficiencies and provide recommendations to ensure a systems safety.

The principles are divided into four categories – “functional specification”, “architectural design”, “implementation methods”, “training, testing and validation” – based on the different aspects of the development process of a system with AI-components.

Guideline for AI in safety-critical applications

Functional specification

1 - Specify intended component and system behavior

For the purpose of verifying and validating the behavior of an AI-based system or component, one needs to clearly define the intended component and system behavior. Consider differences in the specification of the behavior on component and system level.
[4]

2 - Define the Operational Design Domain

To guarantee a safe behavior of AI based components, their field of operation must be clearly specified and bounded. Out of distribution input can cause hazardous behavior of a Neural Network. Determine how the AI-based system shall behave if it is out of the ODD.
[5], [6], [7]

3 - Perform domain specific risk analysis

When considering the development of safety-critical AI applications, a risk assessment is needed to examine the impact on life and limb, property and financial assets, the environment, fundamental rights and ethical principles. For different industries risk assessment shall be based on existing standards. It is important to check whether these standards are designed to accommodate AI technologies.
[1], [2], [3]

Architectural design

4 - Use AI only where necessary
As the use of AI technology already implies a threat to safety, we recommend its usage only where necessary and clearly a superior fit to the task in comparison to a non-AI based solution.
[8], [9]

5 - Implement AI only at component level
The assumptions about stable hierarchical architectures of components, as proposed for example by ISO 26262 for automated driving systems, shall be kept and AI shall solely be used at the component level. This applies equally to non-automotive applications. End-to-end solutions shall be avoided since they are less reliable with regard to verification and validation capabilities. This way, the influence on the overall system behavior can be limited and we encounter AI safety concerns only at the component level, which might be supervised and controlled by a higher instance.
[1], [3], [10], [4]

6 - Prevent too complex AI sub tasks
Define a system architecture assigning component responsibilities based on the overall system requirements. Identify subtasks which are suitable for an AI-based solution. The scope of a single AI-based component shall be appropriate in terms of performance expectation and overall system behavior. If safety requirements are not achievable by the AI unit itself, a solution might be found in the system architecture.
[11], [12]

7 - Make use of redundancy strategies
To increase the robustness of AI components the implementation of redundancy strategies is recommended. A common approach are ensemble strategies. These are classically based on the independent training of multiple classifiers and the subsequent combination of the resulting models based on their performance. Thereby, advantages of different architectures can be combined.
[13], [14], [15], [16]

8 - Employ an independent Safety Monitor
To verify the functionality and safety of the system, an independent supervisor entity shall be implemented. The supervisor entity shall not be logically connected with any other component in terms of the verification of the system. Nevertheless, it must have knowledge about the specified intended functionality of the component, the mandatory data subset to maintain safe system behavior and the capability of verifying against safety standards and requirements. Define measures in case of violation.
[4], [17]

Implementation methods

9 - Ensure decisions are exclusively rule-based
Neural Networks impose a lack of complete specification and deterministic behavior. Rule-based approaches are easier to control in most cases and shall therefore be mandatory for decision making. However, an AI-calculated input can be used as support.

[4], [18], [19]

10 - Select suitable AI approach for each specific task
The specific AI approach, e.g. the architecture of the Neuronal Network, should be well selected for each specific task to ensure a maximum of safety, reliability and suitable performance.

[1], [2], [20], [21]

11 - Base implementation on established coding guidelines
If no coding guidelines for AI-based systems exist in the field of the target application, base your implementation on established coding guidelines for conventional programming.

[22], [23]

Training, Testing and Validation

12 - Define quality gates for training, validation and test data
The training, validation and test data determines the performance and safety of an AI component. It must be ensured that the data represents the realworld application domain, by using quality gates which allow an unbiased evaluation. Top-down technical safety requirements determine mandatory quality gates.
[24], [25]

13 - Define task specific validation metrics
The learning process of a Neural Network is based on the definition of an appropriate set of metrics and a suitable weighting to be optimized for each specific task. In safety-critical application these metrics are derived from Technical Safety Requirements and can for example be the accuracy of a model or its confidence in predictions, which strongly correlates with the safety of the system.
[26], [27], [28], [13]

14 - Apply independent safety assessment
A safety-critical application in which AI is used should be subject to a safety assessment by an independent third party before being launched on the market.
[29], [30], [1]

15 - Develop a maintenance strategy
AI based systems are highly sensible to changes in the input data. A systematic strategy for continuous maintenance is required to mitigate performance decrease by Distributional Shift and its impact on the safety of the system.
[13], [31]

Sources

[1] AutoDrive_SC7_Intermediate_Report, \\ARIANE6.etamax.local\I_internal\I4_ProjectCluster\I41_55_Innovation\I47-074_AutoDrive\Documents\Reports\AutoDrive SC7 Intermediate Re-port\AutoDrive_SC7_Intermediate_Report.pdf.

[2] N. Papernot and I. Goodfellow, "www.cleverhans.io," 14 June 2017. [Online]. Available: www.cleverhans.io/security/privacy/ml/2017/06/14/verification.html. [Accessed January 2022].

[3] R. Salay, R. Queiroz and K. Czarnecki, "arxiv.org," 2017. [Online]. Available: arxiv.org/abs/1709.02435. [Accessed January 2022].

[4] B. Kaiser, "www.researchgate.net," 2017. [Online]. Available: www.researchgate.net/publication/320472118, 2017. [Accessed January 2022].

[5] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman and D. Mane, "arxiv.org," 2016. [Online]. Available: arxiv.org/abs/1606.06565. [Accessed January 2022].

[6] C. W. Lee, D. E. G. Nasif Nayeer, A. Agrawal and B. Liu, "ieeexplore," 08 January 2021. [Online]. Available: ieeexplore.ieee.org/abstract/document/9304552. [Accessed 30 March 2022].

[7] S. Fort, J. Ren and B. Lakshminarayanan, "arxiv.org," 29 July 2021. [Online]. Available: arxiv.org/abs/2106.03004. [Accessed 30 March 2022].

[8] N. Duggal, "simplilearn.com," 03 March 2022. [Online]. Available: www.simplilearn.com/advantages-and-disadvantages-of-artificial-intelligence-article. [Accessed 30 March 2022].

[9] R. Freeman, "towardsdatascience," 24 July 2020. [Online]. Available: towardsdatascience.com/when-to-not-use-ai-or-use-it-based-on-my-experience-abb58c063aba. [Accessed 30 March 2022].

[10] S. M. Grigorescu, M. Glaab and A. Roßbach, "www.elektrobit.com," 2017. [Online]. Available: www.elektrobit.com/newsroom/read-new-techpaper-chances-challenges-using-machine-learning-highly-automated-driving/. [Accessed November 2021].

[11] E. Mueller, "medium.com," 27 May 2021. [Online]. Available: medium.com/capital-one-tech/end-to-end-models-for-complex-ai-tasks-8c34080145cd. [Accessed 30 March 2022].

[12] I. Sydorenko, "labelyourdata.com," 04 May 2021. [Online]. Available: labelyourdata.com/articles/how-to-choose-a-machine-learning-algorithm. [Accessed 30 March 2022].

[13] T. Sämann, P. Schlicht and F. Hüger, "arxiv.org," 2020. [Online]. Available: arxiv.org/abs/2002.08935. [Accessed January 2022].

[14] E. Lutins, "towardsdatascience.com," 2 August 2017. [Online]. Available: towardsdatascience.com/ensemble-methods-in-machine-learning-what-are-they-and-why-use-them-68ec3f9fef5f. [Accessed November 2021].

[15] C. Lundquist, "www.diva-portal.org," 2011. [Online]. Available: www.diva-portal.org/smash/record.jsf. [Accessed January 2022].

[16] J. Kocic, N. Jovicic and V. Drndarevic, "ieeexplore.ieee.org," 2018. [Online]. Available: ieeexplore.ieee.org/abstract/document/8612054. [Accessed January 2022].

[17] M. H. Osman, S. Kugele and S. Shafaei, "ieeexplore.org," 02 January 2020. [Online]. Available: ieeexplore.ieee.org/abstract/document/8945611. [Accessed 30 March 2022].

[18] R. Smith, "becominghuman.ai," 14 July 2020. [Online]. Available: becominghuman.ai/the-key-differences-between-rule-based-ai-and-machine-learning-8792e545e6. [Accessed November 2021].

[19] J. M. Carew, "searchenterpriseai.techtarget.com," 23 July 2020. [Online]. Available: searchenterpriseai.techtarget.com/feature/How-to-choose-between-a-rules-based-vs-machine-learning-system. [Accessed November 2021].

[20] G. D. Luca, "www.baeldung.com," 24 August 2020. [Online]. Available: www.baeldung.com/cs/neural-networks-hidden-layers-criteria. [Accessed November 2021].

[21] A. Tch, "towardsdatascience.com," 14 August 2017. [Online]. Available: towardsdatascience.com/the-mostly-complete-chart-of-neural-networks-explained-3fb6f2367464. [Accessed November 2021].

[22] H. M. Limited, "Misra.org," 2020. [Online]. Available: www.misra.org.uk/app/uploads/2021/06/MISRA-Compliance-2020.pdf. [Accessed 04 April 2022].

[23] M. Barr, "barrgroup.com," 2018. [Online]. Available: barrgroup.com/sites/default/files/barr_c_coding_standard_2018.pdf. [Accessed 05 April 2022].

[24] P. Gupta, "towardsdatascience.com," 5 June 2017. [Online]. Available: towardsdatascience.com/cross-validation-in-machine-learning-72924a69872f. [Accessed November 2021].

[25] W. Koehrsen, "towardsdatascience.com," 28 January 2018 . [Online]. Available: towardsdatascience.com/overfitting-vs-underfitting-a-complete-example-d05dd7e19765. [Accessed November 2021 ].

[26] S. Y. e. al, "arxiv.org," [Online]. Available: arxiv.org/abs/1803.04792.

[27] A. Kumar, "dzone.com," 28 January 2018. [Online]. Available: dzone.com/articles/machine-learning-validation-techniques. [Accessed November 2021 ].

[28] V&V of NN for Aerospace Systems, \\ARIANE6.etamax.local\I_internal\I4_ProjectCluster\I41_55_Innovation\I47-074_AutoDrive\Documents\Reports\AutoDrive SC7 Intermediate Re-port\Report_Relevant_Documents\[21] V&V of NN for Aerospace Systems.pdf.

[29] T. A. LAB, "www.tuev-verband.de," 2021. [Online]. Available: www.tuev-verband.de. [Accessed January 2022].

[30] E. Kommission, "op.europa.eu," 2019. [Online]. Available: op.europa.eu/de/publication-detail/-/publication/d3988569-0434-11ea-8c1f-01aa75ed71a1. [Accessed Januar 2022].

[31] S. H. S. Bhasker, M. Srivastava and K. Yeturu, "doi.org," 2021. [Online]. Available: doi.org/10.36227/techrxiv.14565078.v1. [Accessed January 2022].

[32] J. Brownlee, "machinelearningmastery.com," 6 August 2019. [Online]. Available: machinelearningmastery.com/why-training-a-neural-network-is-hard/. [Accessed November 2021].

[33] T. D. Krafft and K. A. Zweig, "www.vzbv.de," 2019. [Online]. Available: www.vzbv.de/sites/default/files/downloads/2019/05/02/19-01-22_zweig_krafft_transparenz_adm-neu.pdf. [Accessed Januar 2022].

[34] B. J. Taylor, M. A. Darrah and C. D. Moats, "www.spiedigitallibrary.org," 2003. [Online]. Available: www.spiedigitallibrary.org/conference-proceedings-of-spie/5103/0000/Verification-and-validation-of-neural-networks--a-sampling-of/10.1117/12.487527.short. [Accessed January 2022].