Neural Network Integrity Checks Using Golden Sample Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing integrity check mechanisms for neural networks, such as hash operations and checksums, are not practical due to the complexity of neural networks and require system reboots, making them unsuitable for real-time verification during operation.
Innovation Solution
A neural network is trained to recognize a 'golden sample' outside its classification set, producing a preposterous result, allowing for integrity checks without system reboots by comparing the network's output to a predefined probability threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional integrity check mechanisms (hashing, checksum) are applied to neural networks, then data structure integrity can be verified, but system reboot is required which disrupts normal operations
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network with golden samples during the initialization phase. This pre-prepared knowledge enables the network to later detect integrity compromises without requiring system reboot, allowing continuous operation while maintaining verification capability
Solution Approach 2:
The patent introduces golden samples as an intermediary element that mediates between the neural network and integrity verification requirements. These special training samples act as a bridge, enabling the network to self-verify its integrity through normal classification operations without external intervention or system shutdown
2Measurement precision
If neural network weights are adjusted during normal use to improve performance, then classification accuracy improves, but hash or checksum values change making traditional integrity checks invalid
Solution Approach 1:
The patent applies dynamics by making the integrity verification mechanism adaptive to weight changes. Instead of using static hash values that become invalid when weights change, the system dynamically maintains verification capability through golden samples that continue to produce predictable classification outputs regardless of weight adjustments, allowing both performance optimization and integrity monitoring to coexist
Solution Approach 2:
The patent changes the verification parameter from static hash/checksum values to dynamic classification probability vectors. By monitoring whether the neural network's output probabilities for golden samples remain within expected ranges, the system can verify integrity even when other parameters (weights) change during normal operation
3Reliability
If traditional integrity check procedures are implemented, then data structure compromise can be detected, but operations not normally performed by the neural network are required
Solution Approach 1:
The patent applies universality by designing the neural network to perform dual functions: its primary classification task and its secondary integrity verification task. The same neural network architecture and forward propagation mechanism used for normal classification are also used for verifying integrity through golden samples, eliminating the need for separate verification procedures and simplifying operation
Solution Approach 2:
The patent implements self-service by enabling the neural network to verify its own integrity autonomously. During normal operation, the network automatically classifies golden samples and checks whether the output probabilities match expected patterns, allowing self-diagnosis of potential compromises without external verification tools or complex procedural interventions
Data Source
AI summary
A neural network is trained to match digital samples to categories in a set of categories and when presented with at least one golden sample, which is a sample outside the set of categories, to output a probability vector indicative of a preposterous result that the golden sample is matched to a predefined category in the set of categories. The secure computer system is programmed with the trained neural network, adapted to receive digital samples and to present the digital samples to the trained neural network. As an integrity check, the computer system, is caused to present the golden sample to the trained neural network and if the neural network outputs a probability vector classifying the golden sample into a predefined category in a way that is a preposterous result, declaring the neural network as uncompromised and, otherwise, declaring the neural network as compromised.