ANN-based sensitivity analyses

ANN-based sensitivity analysis is a mathematical technique used to systematically evaluate how the output of a trained Artificial Neural Network (ANN) model responds to variations in its input variables. The core of this analysis is 'attribution'—determining which input features have the most significant impact on the model's predictions. In risk management, particularly within automotive functional safety (ISO 26262) and cybersecurity (ISO/SAE 21434), this technique is crucial. For instance, it can quantify how sensitive an autonomous driving perception model is to changes in camera input brightness, contrast, or radar signal noise. According to the 'Measure' function of the NIST AI Risk Management Framework (AI RMF), organizations must assess the robustness and reliability of AI systems. Sensitivity analysis is a key technical means to fulfill this requirement, effectively identifying vulnerabilities that could lead to catastrophic failures and guiding model hardening and risk control design.

In the automotive industry's risk management, this analysis is a critical step for ensuring AI system safety and compliance. The implementation process is as follows: 1. **Model & Parameter Definition**: Select the target ANN model, such as a pedestrian detection model in an ADAS. Identify key input parameters (e.g., image brightness, adverse weather conditions, sensor noise levels) and core output metrics (e.g., detection accuracy, response time). 2. **Systematic Simulation & Testing**: In a Software-in-the-Loop (SIL) or Hardware-in-the-Loop (HIL) environment, systematically vary single or multiple input parameters within predefined ranges and run extensive simulations to record the model's output changes. This process must simulate various real-world edge cases and potential cyber-attack scenarios. 3. **Sensitivity Quantification & Risk Integration**: Use statistical methods (e.g., Sobol indices) to calculate sensitivity scores for each input. High sensitivity to irrelevant noise or environmental changes indicates a potential vulnerability. The findings must be integrated into the Threat Analysis and Risk Assessment (TARA) process required by ISO/SAE 21434 to update risk levels and define corresponding controls. A leading European automaker used this method to reduce false positives in its AEB system caused by sensor noise by 18%, significantly improving product safety.

Taiwanese automotive supply chain companies face three main challenges when implementing this advanced analysis: 1. **High Cost of Integrated Test Platforms**: Building HIL/SIL platforms that can accurately simulate real driving scenarios and cyber-attacks is expensive, posing a financial burden on SMEs. **Solution**: Start with cloud-based simulation services on a pay-as-you-go basis to reduce initial investment and collaborate with institutions like ARTC. 2. **Shortage of Interdisciplinary Talent**: This technique requires professionals with expertise in automotive engineering, AI algorithms, and cybersecurity, a talent pool that is currently limited. **Solution**: Develop internal training programs and partner with universities. Initially, engage external consultants like Winners Consulting for methodology implementation and knowledge transfer. 3. **High Computational Cost**: Comprehensive sensitivity analysis for complex deep learning models demands significant GPU resources, leading to high time and energy costs. **Solution**: Adopt more efficient methods like surrogate model-based analysis to reduce simulations. Prioritize analysis on the highest-risk system modules identified through initial risk assessments to optimize resource allocation.

Winners Consulting specializes in ANN-based sensitivity analyses for Taiwan enterprises, delivering compliant management systems within 90 days. Free consultation: https://winners.com.tw/contact