Traffic Object Detection With Symbolic Checks Against Adversarial Errors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting and classifying traffic light systems in autonomous vehicles using neural networks are prone to errors, especially in rare scenarios, noise, and adversarial attacks, leading to potential safety risks and rule violations.
Innovation Solution
A redundant detection and classification method using both a neural network and a symbolic monitoring algorithm, where the symbolic algorithm provides a comprehensible decision-making process, checks the neural network's output for consistency, and includes a consistency check mechanism to enhance reliability and protect against adversarial attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a neural network is used for detecting and classifying traffic light systems, then object recognition probability is improved, but vulnerability to adversarial attacks and unexpected behavior increases
Solution Approach 1:
A symbolic monitoring algorithm is introduced as an intermediary between the neural network and the control system. This symbolic algorithm verifies the neural network's output by checking against predefined traffic rules and logical constraints, blocking adversarial attacks that would otherwise cause incorrect control decisions.
Solution Approach 2:
The symbolic monitoring algorithm performs preliminary verification of the neural network's detection results before they are used for control decisions. By pre-defining valid traffic scenarios and constraints, the system proactively identifies and rejects adversarial inputs that violate these constraints.
2Measurement precision
If a neural network is used for detecting and classifying traffic light systems, then detection accuracy is improved, but interpretability of decision-making process deteriorates
Solution Approach 1:
The detection system is segmented into two independent modules: a neural network for accurate object detection and classification, and a symbolic monitoring algorithm for interpretable verification. Each module performs its specialized function, with the symbolic part providing explainable checks on the neural network's output.
Solution Approach 2:
The symbolic monitoring algorithm acts as an intermediary that translates the neural network's black-box output into interpretable form by verifying against explicit traffic rules and constraints, making the overall system's decision-making process understandable and explainable.
3Reliability
If redundant detection using both neural network and symbolic monitoring algorithm is implemented, then reliability against adversarial attacks is improved, but system complexity increases
Solution Approach 1:
The symbolic monitoring algorithm is designed as a lightweight intermediary that performs simple logical checks on the neural network's output without requiring complex additional hardware or processing resources. The verification process uses predefined rules that are computationally efficient to evaluate.
Solution Approach 2:
Instead of implementing a fully redundant independent detection system, the patent uses the symbolic monitoring algorithm to create a virtual copy of the verification function that checks the neural network's output against the same traffic rules and constraints, providing redundancy through software-based verification rather than duplicate hardware.
Data Source
AI summary
A method (40) for detecting and classifying at least one object in road traffic is disclosed. Sensor data from a sensor are first of all provided. At least one object is detected and classified on the basis of the sensor data using a neural network. The object is additionally detected and classified on the basis of the sensor data using a symbolic monitoring algorithm. A check is carried out in order to determine whether the neural network and the symbolic monitoring algorithm provide consistent results with respect to the detection and classification of the object.