Diplomarbeit AndreasWeich Absicherung Einsatzstellen Strasse
KI Absicherung
Transcript of KI Absicherung
KI Absicherung
Dr. Stephan Scholz, Volkswagen AG
11th March 2021, Online, Interim presentation
Partner
Consortium Lead:
Volkswagen AG
Co - Lead: Fraunhofer IAIS
Duration: 36 months
01.07.2019 –20.06.2022
24 Partners
Budget: €41M
Funding: €19.2M
External techonolgy partners
2
Project vision and goals
1
Making the safety of AI-based function modules for highly automated driving verifiable
Pedestrian detection
4
Challenge
KI Absicherung Gemeinsame Sprache
Absicherungsmethodik für sicherheitsrelevante KI-Module
Promising new technology withunimagined possibilities
Established safety processes cannotbe applied
Safe, trustworthy driving function
SafetyLandAI Land
Industry consensus (Safe AI): Methodology forjoint safety argumentation
Pixabay Pixabay
KI Absicherung | Interim presentation | 11.03.2021 5
Main goals
6KI Absicherung | Interim presentation | 11.03.2021
KI Absicherung develops and investigates means and methods for verifying AI-based functions for highly
automated driving.
1. Methods for training and testing of AI-based functions
For the pedestrian detection use case, the project is developing an exemplary safety argumentation and
methods for verifying a complex AI function.
2. Safety argumentation
The project‘s results will be used in the exchange with standardization bodies to support the
development of a standard for safeguarding AI-based function modules.
3. Communication with standardization bodies on AI certification
The KI Familie and its projects
KI Absicherung | Interim presentation | 11.03.2021 7
KI Absicherung in connection with the KI Familie
8KI Absicherung | Interim presentation | 11.03.2021
Methodological and conceptual approach
2
From a data-driven AI function to an Assurance Case
10
Data Generation AI FunctionTraining
Improvem
ents Requ
irem
ents
Mitigation effect
Assurance Case Evidence
Methods &Measures MetricsTest
KI Absicherung | Interim presentation | 11.03.2021
Stringent argumentation for the AI function and its Operational Design Domain (ODD).
Development and validation of testing methods for these metrics.
Process-related generation of synthetic learning, testing and validation data.
Development of measures and methods that
improve the AI function over a wide array of
metrics.
Source: BIT Technology Solutions
2354
1Conceptual approach
1. Provide the AI function for pedestrian detection.
2. Generate synthetic learning, testing and validation data.
3. Develop and evaluate measures and methods for the verification of the AI function.
4. Establish an overall safety strategy for the AI function.
5. Define and implement an Assurance Case.
KI Absicherung | Interim presentation | 11.03.2021 11
Project structure with sub-projects and work packages
KI Absicherung | Interim presentation | 11.03.2021 12
Our Approach: Establishment of a Holistic Safety Strategy
KI Absicherung | Interim presentation | 11.03.2021 13Opel
Volkswagen AG)
Identification: Identify potential causes of insufficiencies in the function (in KIA: “DNN related safety concerns”)
Quality metric: Introduce metrics or some form of judgment to argue that insufficiency was mitigated
Countermeasure: Develop methods to mitigate the insufficiencies
Argumentation: Argue that the residual risk associated with the causes has been reduced to a tolerable level
Formalisation: Create the evidence based safety argumentation in a Goal Structuring Notation
Structured safety verification
(Assurance Case)
Specification
KI Absicherung | Interim presentation | 11.03.2021 14
Pre-processingof sensoroutput
CameraPedestrianDetection
(DNN)
Post-processingof the DNN
output
Supervisor/Plausibility
Check
Monitor
Pixabay
Surrounding conditions
System Architecture
Function
Source: Bosch
AI Function-Pedestrian detection
KI Absicherung | Interim presentation | 11.03.2021 15Opel
BMW
Sourrounding conditions
AI Function
System architecture
Function
Semantic Segmentation
Intel
2D Bounding Box Detection
Opel
Instance Segmentation
ZF
3D Bounding Box Detection
BMW
3D Pose estimation
Uni Heidelberg
Modeldesign
Training
ML-Daten
BIT TS
ML-Lifecycle
KI Absicherung | Interim presentation | 11.03.2021 16
[Mackevision]
Volkswagen AG Volkswagen AG
Source: BIT Technology Solutions
Model design
Training
Pre-Processing
ML-Data
Test/ V&V incl. Data
Maintenance
Monitoring
Post-Processing
ML LIFECYCLE
Test/ V&V inkl. Daten
ML-Lifecycle-Validation data
KI Absicherung | Interim presentation | 11.03.2021 17
Absicherungs-
Daten
Mackevision Mackevision
Model design
Training
Pre-Processing
ML-Data
Test/ V&V incl. Daten
Maintenance
Monitoring
Post-Processing
ML LIFECYCLE
DNN-specific safety concern:• Adversarial error: leads to overlooking
(false negative)
Method:• Systematic analysis of adversarial errors • Methods to evaluate adversarial resilience• Counter mechanisms during training or operation
Neurocat
Identify, Measure and & Counteract „DNN-specific „Safety Concerns”
KI Absicherung | Interim presentation | 11.03.2021 18
DNN-specific safety concern:• False positive / negative: Pedestrian detection is
incorrect resp. not robust enough
Method: • Assessment of uncertainty: Stochastic evaluation of a
multitude of model variations (Monte Carlo Dropout)
AI Specific Evidence-Based Safety Argumentation
19
Safety and riskanalysisSafety goal 2Safety goal 2
Safety analysis(HARA, GSA, …)
Safety goals
Hazards / Risks
Safety standards(ISO26262, SOTIF…)
Safety argumentation
Safety requ. 2Safety requ. 2Safetyrequirements
Evidence strategyDNN-Insufficiencies
generalisation
Assurance Case (in GSN)
Assumptions / Context
Safety methods
Best practices
EvidenceSafety
measures
Metrics(Performance
& quality)
DNN specificSafety concerns Data, DNN,and
metrics
AI- based function module
Specification
ODD ground context
(AI-) Function
System architecture
Functionality
Model training
ML-data
Pre-Processing
Monitoring
Test / V&V incl. validation data
ML LIFECYCLE
Model design
Post-Processing
Maintenance
metrics metrics metricsmetrics
Architecturemeasures
DNNmeasures
Teatingmeasures
Data measures
Safety measures & metrics
Parallel Activities – Available for other committeesBringing our results „on the road“
• Establishing contacts with TÜV
• VDTÜV Meeting with the BSI on AI-Safety
• TÜV@Conference 2020 – Meet the Expert
• Collaboration with „Certified AI“
• KI.NRW Initiative driven by Fraunhofer IAIS and the BSI (starting in 2021)
• Projekt Letter of Intent for collaboration
• Contributions to „DIN-Roadmap AI“
• Leading the „Quality and Certification“ Workgroup (Fraunhofer IAIS)
• Contributing to the „Mobility“ Workgroup (BMW)
KI Absicherung | Interim presentation | 11.03.2021 20
Project coordination: Dr. Stephan Scholz, Volkswagen AG
Co-lead and scientific coordination: PD. Dr. Michael Mock, Fraunhofer IAIS
Email: [email protected]