# US 2026/0178907 A1 · Execution behavior analysis text-based ensemble malware detector

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0178907 A1
- Document ID: US20260178907A1
- Status: Published application
- Kind: A1
- Application number: 19/542,587
- Filed: 2026-02-17
- Published: 2026-06-25
- Earliest claimed: 2021-02-10
- Assignees: Palo Alto Networks, Inc.
- Applicants: Palo Alto Networks, Inc.
- Inventors: Sujit Rokka Chhetri; William Redington Hewlett, II
- Web page: https://patentgrep.com/patent/US20260178907A1

## CPC

- G06F 21/561: {Virus type analysis}
- G06N 3/08: Learning methods
- G06F 21/56: Computer malware detection or handling, e.g. anti-virus arrangements
- G06F 21/562: {Static detection}
- G06F 21/566: {Dynamic detection, i.e. detection performed at run-time, e.g. emulation, suspicious activities}
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N 20/20: Ensemble learning
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A malware detector has been designed that uses a combination of NLP techniques on dynamic malware analysis reports for malware classification of files. The malware detector aggregates text-based features identified in different pre-processing pipelines that correspond to different types of properties of a dynamic malware analysis report. From a dynamic malware analysis report, the pre-processing pipelines of the malware detector generate a first feature set based on individual text tokens and a second feature set based on n-grams. The malware detector inputs the first feature set into a trained neural network having an embedding layer. The malware detector then extracts a dense layer from the trained neural network and aggregates the extracted layer with the second feature set to form an input for a trained boosting model. The malware detector inputs the cross-pipeline feature values into the trained boosting model to generate a malware detection output.

## Claims

### Claim 1

A method comprising:
generating, with a machine-learning ensemble that comprises first and second pre-processing pipelines and a trained boosting model, a malware classification from a malware analysis report for a file, wherein generating the malware classification comprises,
processing the malware analysis report in the first pre-processing pipeline which generates a first plurality of values of a first feature set based on individual tokens, wherein the first pre-processing pipeline comprises an embedding layer and a deep neural network and processing the malware analysis report in the first pre-processing pipeline comprises extracting the first plurality of values from a dense layer of the deep neural network that received embeddings of tokens of the malware analysis report from the embedding layer;
processing the malware analysis report in the second pre-processing pipeline which generates a second plurality of values of a second feature set based on token sequences and importance of the token sequences, wherein processing the malware analysis report in the second pre-processing pipeline comprises determining token sequences of the malware analysis report and calculating statistical importance values for the determined token sequences;
generating a feature vector with the first plurality of values and the second plurality of values; and
feeding the feature vector into a trained boosting model to obtain a malware classification for the file.

### Claim 2 (depends on 1)

The method of claim 1, wherein processing in the first pre-processing pipeline further comprises tokenizing the malware analysis report to generate a plurality of tokens and feeding the plurality of tokens into the embedding layer which generates the embeddings.

### Claim 3 (depends on 2)

The method of claim 2, wherein processing in the first pre-processing pipeline further comprises cleaning and filtering the malware analysis report, wherein tokenizing the malware analysis report comprises tokenizing the malware analysis report after the cleaning and filtering.

### Claim 4 (depends on 1)

The method of claim 1, further comprising, prior to generating the malware classification:
determining a token vocabulary with byte-pair encoding run on a training dataset that comprises labeled malware analysis reports; and
initializing the embedding layer with the token vocabulary.

### Claim 5 (depends on 4)

The method of claim 4, further comprising, prior to generating the malware classification, jointly training the deep neural network and the embedding layer.

### Claim 6 (depends on 5)

The method of claim 5, wherein the dense layer was identified as the most informative dense layer of the deep neural network when training a boosting model that yielded the trained boosting model.

### Claim 7 (depends on 1)

The method of claim 1, wherein processing in the second pre-processing pipeline comprises selecting the most informative of the token sequences, wherein the second plurality of values of the second feature set comprises the selected token sequences and the statistical importance values of the selected token sequences.

### Claim 8 (depends on 7)

The method of claim 7, wherein selecting the most informative of the token sequences comprises determining which of the tokens sequences of the malware analysis report correspond to k token sequences learned as most informative from a training dataset of labeled malware analysis reports.

### Claim 9

A non-transitory, machine-readable medium having program code for an ensemble malware detector stored thereon, the program code comprising:
instructions of a first pre-processing pipeline that comprises an embedding layer and a deep neural network, wherein the instructions of the first pre-processing pipeline comprise instructions to extract a first plurality of values of a first feature set from a dense layer of the deep neural network that received embeddings of tokens of a malware analysis report from the embedding layer;
instructions of a second pre-processing pipeline to,
determine token sequences of the malware analysis report;
calculate statistical importance values for the determined token sequences; and
determine a second plurality of values of a second feature set based, at least in part, on the determined token sequences and their statistical importance values; and
instructions to generate a feature vector with the first plurality of values and the second plurality of values and feed the feature vector into a trained boosting model to obtain a malware classification for a file corresponding to the malware analysis report.

### Claim 10 (depends on 9)

The non-transitory, machine-readable medium of claim 9, wherein the instructions of the first pre-processing pipeline further comprise instructions to tokenize the malware analysis report to generate a plurality of tokens and feed the plurality of tokens into the embedding layer which generates the embeddings.

### Claim 11 (depends on 10)

The non-transitory, machine-readable medium of claim 10, wherein the program code further comprises instructions to clean and filter the malware analysis report prior to tokenization.

### Claim 12 (depends on 9)

The non-transitory, machine-readable medium of claim 9, wherein the dense layer was identified as the most informative dense layer of the deep neural network when training a boosting model that yielded the trained boosting model.

### Claim 13 (depends on 9)

The non-transitory, machine-readable medium of claim 9, wherein the instructions of the second pre-processing pipeline comprise instructions to select the most informative of the determined token sequences, wherein the second plurality of values of the second feature set comprises the selected token sequences and the statistical importance values of the selected token sequences.

### Claim 14 (depends on 13)

The non-transitory, machine-readable medium of claim 13, wherein the instructions of the second pre-processing pipeline to select the most informative of the token sequences comprise instructions to determine which of the tokens sequences of the malware analysis report correspond to k token sequences learned as most informative from a training dataset of labeled malware analysis reports.

### Claim 15

An apparatus comprising:
a processor; and
a machine-readable medium having stored thereon instructions for an ensemble malware detector that comprises an embedding layer, a deep neural network, and a trained boosting model, the instructions executable by the processor to cause the apparatus to,
extract a first plurality of values of a first feature set from a dense layer of the deep neural network that received embeddings of tokens of a malware analysis report from the embedding layer;
determine token sequences of the malware analysis report;
calculate statistical importance values for the determined token sequences;
determine a second plurality of values of a second feature set based, at least in part, on the determined token sequences and their statistical importance values; and
generate a feature vector with the first plurality of values and the second plurality of values and feed the feature vector into a trained boosting model to obtain a malware classification for a file corresponding to the malware analysis report.

### Claim 16 (depends on 15)

The apparatus of claim 15, wherein the machine-readable medium further has stored thereon instructions executable by the processor to cause the apparatus to tokenize the malware analysis report to generate a plurality of tokens and feed the plurality of tokens into the embedding layer which generates the embeddings.

### Claim 17 (depends on 16)

The apparatus of claim 16, wherein the machine-readable medium further has stored thereon instructions executable by the processor to cause the apparatus to clean and filter the malware analysis report prior to tokenization.

### Claim 18 (depends on 15)

The apparatus of claim 15, wherein the dense layer was identified as the most informative dense layer of the deep neural network when training a boosting model that yielded the trained boosting model.

### Claim 19 (depends on 15)

The apparatus of claim 15, wherein the instructions to determine the second plurality of values of the second feature comprise instructions executable by the processor to cause the apparatus to select the most informative of the determined token sequences, wherein the second plurality of values of the second feature set comprises the selected token sequences and the statistical importance values of the selected token sequences.

### Claim 20 (depends on 19)

The apparatus of claim 19, wherein the instructions to select the most informative of the token sequences comprise instructions executable by the processor to cause the apparatus to determine which of the tokens sequences of the malware analysis report correspond to k token sequences learned as most informative from a training dataset of labeled malware analysis reports.
