Diffusion-Based Outlier Synthesis for Anomaly Detection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anomaly detection models face challenges in training and evaluation due to the lack of labeled anomalies, leading to biased performance estimation and under-explored evaluation methods, especially when deploying in new applications with unknown outlier distributions.
Innovation Solution
A system and method for anomaly detection using diffusion-based outlier synthesis, generating synthetic outliers through diffusion-based generation techniques, allowing for training and validation of anomaly detection models in environments with limited labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If anomaly detection models are trained without sufficient labeled anomaly data, then training can proceed with available normal data, but model performance in detecting critical anomalies deteriorates
Solution Approach 1:
The patent generates synthetic anomaly data by copying and transforming normal data through diffusion models. The system creates artificial anomaly samples that mimic real anomaly characteristics, enabling training without requiring actual labeled anomaly data. This copying approach allows the model to learn anomaly patterns from synthesized examples rather than requiring scarce real anomaly copies.
Solution Approach 2:
The patent applies parameter changes by modifying the diffusion process parameters to generate diverse anomaly variations. By adjusting guidance strength, noise schedules, and transformation parameters, the system creates multiple anomaly types from normal data, expanding the training dataset with parameter-varied synthetic anomalies that improve detection reliability.
2Productivity
If evaluation is performed without sufficient labeled abnormal data, then evaluation can be conducted with available resources, but performance estimation becomes biased
Solution Approach 1:
The patent performs preliminary action by generating synthetic anomaly data before evaluation. The system pre-generates a comprehensive set of labeled anomaly samples using diffusion models, then uses these pre-prepared synthetic anomalies for model evaluation. This preliminary data preparation ensures that evaluation can proceed with sufficient test cases, eliminating bias from data scarcity.
Solution Approach 2:
The patent uses copying to create synthetic test sets that replicate the distribution and characteristics of real anomalies. By copying normal data through diffusion transformations to generate synthetic anomaly test samples, the system creates an evaluation dataset that accurately reflects real-world anomaly patterns, enabling unbiased performance measurement.
3Quantity of substance
If diffusion-based generation is used to create synthetic outliers, then training data availability improves, but computational complexity increases
Solution Approach 1:
The patent applies partial action by generating only the necessary number of synthetic anomaly samples required for effective training, rather than exhaustively generating all possible variations. The system uses targeted diffusion generations with controlled guidance strength to produce sufficient anomaly data without unnecessary computational overhead, balancing data quantity with computational feasibility.
Data Source
AI summary
A computer-implemented system and method relate to anomaly detection. Latent code of a source image is obtained. The latent code is designated as a target image. Source embedding data is generated form the source image. Text data, which is of a different domain than that of the source image, is obtained. Text embedding data is generated from the text data. Additional embedding data is generated using the source embedding data and the text embedding data. The additional embedding data provides guidance for modifying the source image. A modified image is generated via an iterative process that includes at least one iteration, where each iteration includes at least (i) encoding the target image to generate target embedding data, (ii) generating updated embedding data by combining the target embedding data and the additional embedding data, (iii) decoding the updated embedding data to generate a new image, and (iv) assigning the new image as the target image and the modified image. A non-anomalous label is generated for the source image and an anomalous label is generated for the modified image. A machine learning model is trained or fine-tuned using a dataset, which includes at least the source image with the non-anomalous label and the modified image with the anomalous label.