Neural Network Latent Space Verification for Adversarial Attack Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models are vulnerable to adversarial attacks, where imperceptible noise can lead to incorrect recognition, posing a risk to inference systems in applications like image recognition and autonomous vehicles.
Innovation Solution
An information processing system utilizing a neural network model with a latent space database to infer input data based on the positional relationship between output vectors of a hidden layer, enhancing robustness against adversarial attacks by leveraging the positional information in the latent space.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network model is used for inference, then high performance in image recognition and classification is achieved, but the system becomes vulnerable to adversarial attacks with imperceptible noise causing incorrect recognition
Solution Approach 1:
The patent introduces an intermediary verification mechanism that compares the neural network's output with reference information about the input data. This intermediary layer acts as a mediator between the neural network's potentially flawed output and the final decision, checking whether the output is consistent with expected patterns and characteristics of the input data, thereby detecting adversarial attacks without compromising the neural network's recognition capabilities
Solution Approach 2:
The patent implements a feedback mechanism where the system continuously monitors the consistency between neural network outputs and reference information. When inconsistencies are detected that suggest adversarial manipulation, the system can adjust its processing or reject the output, creating a closed-loop control system that improves reliability while maintaining recognition performance
2Object-affected harmful factors
If noise is added to input data by malicious attackers, then incorrect recognition is induced in the neural network, but the noise remains imperceptible to human eyesight
Solution Approach 1:
The patent replaces reliance on human visual perception with computational verification mechanisms. Instead of depending on human eyesight to detect noise (which cannot perceive adversarial perturbations), the system uses automated consistency checks between neural network outputs and reference information, substituting mechanical/human detection with computational analysis that can identify subtle adversarial manipulations
Solution Approach 2:
The patent creates reference copies or representations of expected input data characteristics and compares them against actual inputs. By having a copy or model of what legitimate input data should look like, the system can detect deviations caused by adversarial noise without requiring human perception, using the reference copy as a benchmark for validation
Data Source
AI summary
To provide a robust information processing system against attacks by Adversarial Example. A neural network model 608, a latent space database 609 for storing position information in a latent space in which first output vectors, which are output vectors of a predetermined hidden layer included in the neural network model, are embedded concerning input data used for learning of the neural network model, and an inference control unit 606 for making an inference using the neural network model and the latent space database are provided. The inference control unit infers the input data based on the positional relationship between the second output vector, which is an output vector of the predetermined hidden layer concerning input data to be inferred, and the first output vectors in said latent space.