Encoder-Decoder Neural Network for Application Log Event Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for converting character-based application log files into numerical formats for machine-learning systems are inadequate, failing to provide sufficient quality for optimal performance in detecting application events such as digital attacks and data theft.

Innovation Solution

An encoder-decoder neural network is trained to encode and decode application log values, optimizing a metric to generate high-quality collectively encoded values that enable detection of application events by clustering and identifying outliers in machine-learning systems.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If simple conversion methods are used to convert character values into numbers, then the conversion process is easy and fast, but the quality of conversion is inadequate for optimal machine-learning system performance

Engineering Contradiction:

Improveconversion qualityVSAvoidconversion method complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

An encoder-decoder neural network is introduced as an intermediary system between the character-based log files and the machine-learning system. The encoder converts character values to numerical embeddings, and the decoder reconstructs the original values, ensuring high conversion quality while maintaining automated processing

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The conversion process transforms character values into numerical embeddings by changing the parameter representation from discrete characters to continuous numerical vectors. This parameter transformation enables the machine-learning system to process log data with high precision while the neural network automatically optimizes the conversion quality

Inventive Principle:

Principle #35Parameter changes

2Reliability

If character-based log files are processed directly without proper encoding, then the processing is simpler, but the machine-learning system cannot effectively detect application events

Engineering Contradiction:

Improvedetection accuracyVSAvoidencoding system complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The traditional mechanical conversion methods are replaced with a neural network-based encoding system. The encoder-decoder neural network automatically learns the optimal transformation from characters to numerical representations, significantly improving detection reliability for application events while the automated neural network processing maintains efficiency

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Productivity

If complex encoding methods are used to improve conversion quality, then machine-learning system performance improves, but the encoding process becomes more complex and resource-intensive

Engineering Contradiction:

Improvemachine-learning system performanceVSAvoidencoding process complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The encoder-decoder neural network is trained in advance on representative log data to learn optimal encoding transformations. This preliminary training action enables the encoder to efficiently convert new log data with high quality without requiring complex real-time processing, thereby improving machine-learning system performance while maintaining encoding efficiency

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The neural network creates numerical embedding copies of the character-based log data that preserve the essential information needed for machine-learning analysis. These numerical copies enable high-performance processing by the machine-learning system while the encoding process itself remains automated and efficient

Inventive Principle:

Principle #26Copying

Data Source

PatentUS11567850B2Detecting application events based on encoding application log values

Publication Date: 2023.01.31 SALESFORCE INC

AI summary

An encoder receives an application log file including component values and encodes the component values into lists of preliminary encoded values. The lists of preliminary encoded values are combined into a combined list of preliminary encoded values. An encoder-decoder neural network is trained to encode the combined list of preliminary encoded values into a list of collectively encoded values, to decode the list of collectively encoded values into a list of decoded values, and to optimize a metric measuring the encoder-decoder neural network's functioning, in response to receiving the combined list of preliminary encoded values. The trained encoder-decoder neural network receives combined lists of preliminary encoded values for application log files and encodes the combined lists of preliminary encoded values into lists of collectively encoded values. The lists of collectively encoded values are sent to a detector, thereby enabling the detector to detect an application event associated with the application log files.