Classification Result Verification Using Clustering for Edge AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing detection schemes for adversarial examples in machine learning/deep learning models require significant computational resources, making them impractical for edge devices and leading to high latency, which is problematic for real-time applications like autonomous driving.
Innovation Solution
A method involving clustering feature representations of training samples using true-value labels to determine a clustering class, verifying the reliability of classification results based on this class, and integrating domain-specific knowledge to improve detection precision while reducing computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing detection schemes for adversarial examples are implemented, then detection precision is improved, but computational resources required increase significantly
Solution Approach 1:
The patent divides the detection process into two independent components: a lightweight verification model that runs on edge devices for preliminary filtering, and a more comprehensive detection scheme that operates on centralized servers. This segmentation allows edge devices to perform low-latency verification with minimal computational resources while maintaining detection precision through the coordinated operation of both components.
Solution Approach 2:
The verification model acts as an intermediary between edge devices and the centralized detection system. It performs preliminary verification of classification results using domain-specific knowledge and clustering information, filtering out obvious adversarial examples before they reach the resource-intensive centralized detection scheme, thus reducing overall computational resource consumption.
2Measurement precision
If existing detection schemes for adversarial examples are implemented, then detection precision is improved, but latency increases
Solution Approach 1:
By segmenting the detection system into edge-based verification and centralized detection, the patent enables real-time verification at the edge with low latency, while comprehensive detection occurs centrally without blocking real-time operations. This resolves the contradiction between high detection precision and low latency.
Solution Approach 2:
The verification model performs preliminary verification of classification results at the edge device before results are finalized. This preliminary action uses pre-computed clustering information and domain-specific knowledge to quickly identify and filter adversarial examples, reducing the need for time-consuming comprehensive detection in real-time applications.
3Reliability
If verification based on clustering classes is performed, then reliability of classification results is improved, but device complexity increases
Solution Approach 1:
Clustering information and domain-specific knowledge are pre-computed during the training phase and stored in the verification model. During inference, the verification process simply compares classification results against these pre-computed references, maintaining high reliability while avoiding the complexity of performing complex clustering computations in real-time on edge devices.
Solution Approach 2:
The verification model uses pre-stored clustering information and domain-specific knowledge to perform self-verification of classification results without requiring complex external resources or computations. This self-service approach maintains reliability while keeping the verification process simple and lightweight for edge device deployment.
Data Source
AI summary
Embodiments of the present disclosure relate to a method, a device, and a computer program product for verifying a classification result. The method comprises: obtaining a plurality of clusters for training samples by clustering feature representations of the training samples using labels of the training samples; determining, based on the plurality of clusters, a clustering class to which an input image belongs; acquiring a classification result of a classification model for the input image; and verifying the degree of reliability of the classification result for the input image based on the clustering class and the classification result. In this manner, the verification according to embodiments of the present disclosure not only easily combines domain-specific knowledge and improves the detection precision, but also saves computational overhead and storage resources, thus enabling the solution to be deployed in edge devices or Internet of Things devices with limited computational power.