Multi-Model Anomaly Detection for Neural Network Adversarial Defense
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI systems are vulnerable to cybersecurity threats such as adversarial attacks, where malicious actors subtly modify input data to deceive AI models, leading to compromised system integrity, especially in critical applications like healthcare.
Innovation Solution
A multi-model approach is implemented at the input level of AI systems, using three models each with three distinct tests to detect anomalies. This approach employs a 2oo3 voting scheme, where data is filtered if at least two out of three models identify it as anomalous, thereby reducing bias and enhancing detection reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single anomaly detection test is used in the AI system, then the device complexity is low, but the reliability of anomaly detection is insufficient and biased results may occur
Solution Approach 1:
The anomaly detection system is segmented into three distinct models (first model, second model, third model), each performing specific tests on input data. This segmentation allows each model to specialize in detecting different types of anomalies while maintaining overall system reliability through distributed detection responsibilities.
Solution Approach 2:
The patent combines multiple anomaly detection models into a unified system where their results are integrated through a voting mechanism. The models work together as an ensemble, merging their individual detection capabilities to achieve more reliable and balanced anomaly detection than any single model could provide alone.
2Measurement precision
If multiple models with unique signatures are used to detect anomalies, then detection reliability and bias reduction are improved, but the device complexity increases
Solution Approach 1:
Each anomaly detection model is assigned a unique signature or specialization, giving it local quality in detecting specific types of anomalies. The first model, second model, and third model each have distinct detection capabilities tailored to different anomaly patterns, improving overall detection precision through specialized local expertise.
Solution Approach 2:
The processing system is designed with universal capability to execute multiple different anomaly detection models with various signatures. This multi-functional processing system can adaptively apply different detection strategies depending on the input data characteristics, achieving high measurement precision across diverse anomaly types.
3Reliability
If a 2oo3 voting scheme is implemented to filter anomalous data, then the integrity of AI systems is enhanced, but the processing time and complexity increase
Solution Approach 1:
The system performs preliminary anomaly detection using multiple models before final data processing and AI execution. By conducting anomaly detection in advance through the 2oo3 voting scheme, potentially harmful data is identified and filtered out before it can impact the AI system, preventing integrity issues rather than correcting them later.
Solution Approach 2:
The voting scheme extracts and removes anomalous data points from the input dataset based on the 2oo3 detection criterion. Data that is flagged as anomalous by at least two out of three models is extracted from the valid data set and excluded from further AI processing, ensuring system integrity while maintaining efficient processing of clean data.
Data Source
AI summary
A system includes a machine learning (ML) engine in which a data store is coupled to a processing system. The processing system executes code to receive a dataset from the data store, and produces three models each comprising three tests. Each of the tests seeks to detect one of three anomaly types corresponding to each of the models. The processing system performs at least two of the three tests relating to each of the three anomaly types. Separately for each anomaly type, the processing system detects an anomaly when two-out-of-three (2oo3) tests conclude that the anomaly is present in the dataset. The dataset including the flagged anomalies is stored in a data repository. The anomaly is filtered from the dataset. The processing system is configured to use data from the dataset to retrain an existing trained ML model.