Neural Network Malicious File Detection Using Single Sample Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for identifying malicious files, especially in cases of targeted attacks, are ineffective when only a single malicious file is available, as they struggle to detect similar files or those with indirect malicious attributes, and are often fooled by files signed with trusted certificates performing seemingly harmless actions.

Innovation Solution

A learning model trained on a malicious file using a neural network is generated by selecting and modifying parameters such as the number of layers, neurons, and loss function, and then used to identify unknown files by determining if they fulfill detection rules based on attribute vectors, with the ability to retrain the model upon false alarms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional signature and heuristic analysis methods are used to identify malicious files, then files with known signatures can be detected, but files with indirect malicious attributes or signed with trusted certificates cannot be effectively identified

Engineering Contradiction:
Improvemalicious file detection accuracyVSAvoidability to detect novel malicious patterns
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary analysis by extracting multiple attributes from the malicious file sample before training the neural network. This preliminary action includes collecting file headers, metadata, code structure, and behavioral characteristics, which are then used as training data to enable the model to detect similar patterns in unknown files

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The neural network learns to copy the malicious patterns from the training file by creating a mathematical model that replicates the harmful behavior characteristics. This copied knowledge allows the system to identify files with similar malicious attributes without needing the actual malicious file present

Inventive Principle:
Principle #26Copying

2Loss of information

If virtual machine sandboxing is used to analyze file behavior, then detailed behavioral data can be collected, but computational resources and analysis time are significantly increased

Engineering Contradiction:
Improvebehavioral data collection completenessVSAvoidanalysis speed
Core Design Contradiction:
Loss of informationVSProductivity

Solution Approach 1:

The system extracts only the most relevant attributes from the malicious file, such as file headers, metadata, code structure, and key behavioral characteristics. This extraction approach eliminates unnecessary data collection steps while retaining the essential information needed for accurate detection

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system changes the analysis parameters by training the neural network on a comprehensive set of file attributes including binary data, text content, metadata, and code structure. This parameter transformation enables the model to make rapid predictions without requiring actual file execution in a sandbox environment

Inventive Principle:
Principle #35Parameter changes

3Reliability

If neural networks are trained on multiple malicious files, then detection coverage improves, but when only a single malicious file is available, the system cannot effectively identify similar files

Engineering Contradiction:
Improvedetection coverageVSAvoidnumber of training samples
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The system performs preliminary attribute extraction and feature engineering on the single available malicious file, creating a comprehensive training dataset from limited resources. This includes extracting multiple file representations and generating synthetic training examples to maximize the utility of the single sample

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The neural network is designed to learn universal malicious patterns that can be applied across different file types and variations. By focusing on fundamental malicious behavior characteristics rather than file-specific details, the model achieves broad detection coverage from a single training example

Inventive Principle:
Principle #6Universality (Multi-functionality)

4Ease of operation

If files signed with trusted certificates are analyzed using traditional methods, then legitimate files are not flagged, but malicious files with trusted certificates are incorrectly classified as safe

Engineering Contradiction:
Improvefalse positive rateVSAvoidtrusted certificate validation
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The system segments the file analysis into multiple independent attribute evaluations, separating the certificate validation from the malicious behavior detection. This allows the neural network to evaluate file attributes independently of the certificate status, preventing trusted certificates from masking malicious behavior

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of trusting files with valid certificates by default, the system inverts the approach by actively searching for malicious attributes regardless of certificate status. The neural network is trained to identify malicious patterns even in files with trusted certificates, reversing the traditional assumption that valid signatures guarantee safety

Inventive Principle:
Principle #13The other way round (Inversion)

Data Source

PatentEP3588351B1System and method of identifying malicious files using a learning model trained on a malicious file
Publication Date: 2021.04.21 AO KASPERSKY LAB
  • EP3588351B1 patent drawingFigure 1
  • EP3588351B1 patent drawingFigure 2
  • EP3588351B1 patent drawingFigure 3

AI summary

Disclosed herein are systems and methods of identifying malicious files using a learning model trained on a malicious file. In one example, an exemplary method comprises selecting, using a hardware processor, the malicious file from a plurality of malicious files that are known to be harmful, selecting, using the hardware processor, a plurality of safe files from a set of safe files that are known to be safe, generating, using the hardware processor, a learning model by training a neural network with the malicious file and the plurality of safe files, generating, using the hardware processor, rules for detection of malicious files from the learning model, determining, using the hardware processor, whether attributes of an unknown file fulfill the rules for detection of malicious files using the learning model and responsive to determining that the rules for detection are fulfilled, identifying, using the hardware processor, the unknown file as malicious.