In-Vehicle AI Verification Using Context-Based Robustness Testing

The patent introduces an intermediary explanation layer between the deep neural network and the user. This layer generates human-interpretable explanations for the network's decisions, allowing the complex processing capability of deep neural networks to be maintained while restoring interpretability through a mediating component that translates internal representations into understandable forms.

The patent applies preliminary action by incorporating robustness considerations into the training process before deployment. Adversarial training is performed during the development phase, where the system is pre-exposed to perturbed inputs and learns to generalize better, thereby maintaining development efficiency while improving robustness against future adversarial attacks.

The patent employs parameter changes by systematically varying the parameters of synthetic data generation to better match real-world conditions. By adjusting parameters such as noise characteristics, lighting conditions, and object properties in the simulation, the synthetic training data becomes more representative of real sensor data, thereby maintaining the efficiency benefits of simulation while improving transfer performance.

The patent applies segmentation by dividing the infinite space of possible disturbances into discrete, manageable categories and types. By segmenting disturbances into classes such as sensor noise, weather conditions, adversarial attacks, and occlusions, the system can systematically test and verify robustness against each category without being overwhelmed by the infinite complexity of all possible disturbances.