Anomaly Detection via Continuous and Quantized Model Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection systems face challenges in efficiently managing computing resources and ensuring data security while detecting anomalies and data drift, as continuous inference models may erroneously label non-anomalous data as anomalous and adapt to data drift, leading to potential misclassification and security vulnerabilities.
Innovation Solution
The use of both continuous and quantized inference models, where the continuous model detects anomalies and the quantized model is less sensitive to data drift, allowing for comparison of their outputs to identify data drift and enabling adaptive re-training to maintain accurate anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a continuous inference model is used for anomaly detection, then the model can adapt to new data patterns, but it may erroneously label non-anomalous data as anomalous and adapt to data drift
Solution Approach 1:
The system segments the inference process into two distinct models: a continuous inference model for adaptability and a quantized inference model for stability. Each model serves a specific function, with the continuous model capturing data drift and the quantized model providing reliable anomaly detection benchmarks, resolving the contradiction between adaptability and reliability.
Solution Approach 2:
The quantized inference model acts as an intermediary reference that mediates between the continuous model's adaptability and the need for reliable anomaly detection. By comparing outputs from both models, the system identifies true anomalies while filtering out false positives caused by data drift adaptation.
2Measurement precision
If the continuous inference model is re-trained to expand its ability to detect non-anomalous data, then detection accuracy improves, but the model adapts to data drift causing misclassification
Solution Approach 1:
The system implements dynamic model operation where the continuous inference model is re-trained to improve precision, while the quantized inference model remains static to maintain classification reliability. This dynamic approach allows the system to benefit from improved detection precision without sacrificing classification reliability.
Solution Approach 2:
The system changes the parameter state of the inference models by using continuous precision for the adaptive model and quantized discrete states for the reference model. This parameter differentiation allows the continuous model to improve detection precision through re-training while the quantized model maintains classification reliability through its discrete, stable parameter representation.
3Measurement precision
If both continuous and quantized inference models are used, then data drift detection accuracy improves, but computing resource expenditure increases
Solution Approach 1:
The system creates a simplified copy of the continuous inference model in quantized form. This quantized copy requires fewer computing resources while maintaining the essential functionality for anomaly detection, allowing the system to run both models with reduced overall resource expenditure compared to running only high-precision continuous models.
Data Source
AI summary
Methods and systems for detecting data drift while performing anomaly detection in a distributed environment are disclosed. To perform anomaly detection, a system may include an anomaly detector and one or more data collectors. The anomaly detector may detect anomalies in data obtained from one or more of the data collectors using a continuous inference model. To detect data drifts in data from the one or more data collectors, the anomaly detector may also detect anomalies in data obtained from one or more data detectors using a quantized inference model. The output of the continuous inference model may be compared to the output of the quantized inference model to determine whether the continuous inference model has adapted to data drift over time through re-training. Following anomaly detection and/or data drift detection, the data may be discarded to remove the data from the anomaly detector.