Granted Patent B2

US 12,737,461 · App. 18/593,851 · Granted Sep 15, 2026

Inventors: Li Kang (Sammamish, WA); Heon Joon Choi (Seattle, WA); Shauharda Khadka (Seattle, WA); Ankit Garg (Seattle, WA); Jayronn Christian Sulit Bucu (Bellevue, WA); Miranda Nicole McClellan (Dallas, TX); Karen Lavi (Seattle, WA); Siddharth Pavithran (Redmond, WA); Niveditha Shenoy Badiadka (Bothell, WA); Srishti Agarwal (Sunnyvale, CA); Tanmay Arun Ganacharya (Kirkland, WA); Geoffrey Lyall McDonald (Vancouver, CA)

Assignee: Microsoft Technology Licensing, LLC

G06F21/56G06N5/04G06N20/00

Abstract

Claims (113)

1 . A system comprising:

a processor system; and

a memory that stores computer-executable instructions that are executable by the processor system to at least:

generate a target vector based at least on an identifier that identifies a target file, content of the target file, and an AI prompt;

determine extents to which historical vectors associated with historical files, which are previously classified by an AI model, correspond to the target vector;

select a first historical vector from the historical vectors based at least on the extent to which the first historical vector corresponds to the target vector being greater than or equal to an extent threshold;

cause the AI model to classify the target file in a malicious class or a non-malicious class and to provide an explanation that describes reasoning used by the AI model in choosing the malicious class or the non-malicious class by providing the AI prompt together with the target file and classification information associated with the first historical vector as inputs to the AI model, the AI prompt requesting that the AI model choose between the malicious class and the non-malicious class and further requesting the explanation that describes the reasoning used by the AI model in choosing the malicious class or the non-malicious class;

receive a response to the AI prompt from the AI model, the response comprising an indication of a selected class in which the AI model classifies the target file and further comprising the explanation that describes the reasoning used by the AI model in choosing the selected class, wherein the selected class is the malicious class or the non-malicious class;

provide the explanation, which describes the reasoning used by the AI model in choosing the selected class, to a user via a user interface; and

based at least on the indication of the selected class in which the AI model classifies the target file, perform a security operation with regard to the target file.

2 . The system of claim 1 , wherein the computer-executable instructions are executable by the processor system to at least:

cause the AI model to classify the target file in the malicious class or the non-malicious class by providing the AI prompt together with the target file and contextual information regarding the target file as inputs to the AI model; and

wherein the contextual information comprises at least one of the following:

an indication of a location from which the target file is downloaded;

a process path associated with the target file;

a process command line associated with the target file;

a parent process command line associated with the target file; or

a real time process name associated with the target file.

3 . The system of claim 1 , wherein the explanation in the response to the AI prompt indicates a plurality of attributes of the target file that contributed to the AI model choosing the selected class.

4 . The system of claim 1 , wherein the computer-executable instructions are executable by the processor system to at least:

store classification information in a store,

wherein the classification information comprises a reference identifier that identifies a reference file, reference content of the reference file, a reference AI prompt requesting that the AI model choose between the malicious class and the non-malicious class and further requesting a reference explanation that describes reference reasoning used by the AI model in choosing the malicious class or the non-malicious class, and a reference response to the reference AI prompt,

wherein the reference response comprises an indication of a reference class in which the reference file is classified and further comprises the reference explanation that describes the reference reasoning for classifying the reference file in the reference class, wherein the reference class is the malicious class or the non-malicious class;

in response to the classification information being stored in the store, determine that the target file is to be classified among a plurality of classes that comprises the malicious class and the non-malicious class; and

cause the AI model to classify the target file in the malicious class or the non-malicious class and to provide the explanation that describes the reasoning used by the AI model in choosing the malicious class or the non-malicious class by providing the AI prompt together with the target file and the classification information to the AI model.

5 . The system of claim 4 , wherein the classification information further comprises a confidence indicator that indicates a confidence associated with a classification of the reference file in the malicious class or the non-malicious class.

6 . The system of claim 4 , wherein the classification information further comprises a second reference identifier that identifies a second reference file, second reference content of the second reference file, a second reference AI prompt requesting that the AI model choose between the malicious class and the non-malicious class and further requesting a second reference explanation that describes second reference reasoning used by the AI model in choosing the malicious class or the non-malicious class, and a second reference response to the second reference AI prompt; and

wherein the second reference response comprises an indication of a second reference class in which the second reference file is classified and further comprises the second reference explanation that describes the second reference reasoning for classifying the second reference file in the second reference class, wherein the second reference class is the malicious class or the non-malicious class.

7 . The system of claim 1 ,

wherein the historical vectors are based at least on identifiers that identify the historical files, contents of the historical files, and AI prompts requesting that the AI model choose between the malicious class and the non-malicious class and further requesting explanations that describe reasonings used by the AI model in choosing the malicious class or the non-malicious class;

wherein the classification information comprises a first identifier that identifies a first historical file of the historical files that is associated with the first historical vector, first content of the first historical file, a first AI prompt requesting that the AI model choose between the malicious class and the non-malicious class and further requesting a first explanation that describes a first reasoning used by the AI model in choosing the malicious class or the non-malicious class, and a first response to the first AI prompt from the AI model, and

wherein the first response comprises an indication of a first class in which the AI model classifies the first historical file and further comprises the first explanation that describes the first reasoning used by the AI model in choosing the first class, wherein the first class is the malicious class or the non-malicious class.

8 . The system of claim 7 , wherein the historical vectors are further based at least on confidences associated with classifications of the historical files in the malicious class or the non-malicious class; and

wherein the classification information further comprises a first confidence indicator that indicates a first confidence of the confidences that is associated with a first classification of the first historical file in the malicious class or the non-malicious class.

9 . The system of claim 7 , wherein the computer-executable instructions are executable by the processor system further to at least:

select a second historical vector from the historical vectors based at least on the extent to which the second historical vector corresponds to the target vector being greater than or equal to the extent threshold,

wherein the classification information further comprises a second identifier that identifies a second historical file of the historical files that is associated with the second historical vector, second content of the second historical file, a second AI prompt requesting that the AI model choose between the malicious class and the non-malicious class and further requesting a second explanation that describes a second reasoning used by the AI model in choosing the malicious class or the non-malicious class, and a second response to the second AI prompt from the AI model, and

wherein the second response comprises an indication of a second class in which the AI model classifies the second historical file and further comprises the second explanation that describes the second reasoning used by the AI model in choosing the second class, wherein the second class is the malicious class or the non-malicious class.

10 . The system of claim 1 ,

wherein the historical vectors are based at least on identifiers that identify the historical files, contents of the historical files, and AI prompts requesting that the AI model choose between the malicious class and the non-malicious class and further requesting explanations that describe the reasonings used by the AI model in choosing the malicious class or the non-malicious class;

wherein the computer-executable instructions are executable by the processor system to at least:

select N first historical vectors, which are associated with historical files that are classified in the malicious class, and N second historical vectors, which are associated with historical files that are classified in the non-malicious class, from the historical vectors to provide 2N sample historical vectors based at least on the extents to which the N first historical vectors and the N second historical vectors correspond to the target vector being greater than or equal to an extent threshold, wherein N is a positive integer; and

select P sample historical vectors from the 2N sample historical vectors, wherein P is a positive integer that is less than 2N;

wherein the classification information comprises P identifiers that identify P historical files of the historical files that are associated with the P sample historical vectors, contents of the P historical files, P AI prompts requesting that the AI model choose between the malicious class and the non-malicious class and further requesting explanations that describe reasonings used by the AI model in choosing the malicious class or the non-malicious class, and P responses to the P AI prompts from the AI model; and

wherein the P responses comprise indications of selected classes in which the AI model classifies the P historical files and further comprise the explanations that describe the reasonings used by the AI model in choosing the selected classes, wherein a selected class is the malicious class or the non-malicious class.

11 . The system of claim 1 , wherein the computer-executable instructions are executable by the processor system to at least:

cause the AI model to classify the target file, to provide the explanation, and to provide an indication of a confidence associated with a classification of the target file in the malicious class or the non-malicious class,

wherein the AI prompt further requests the indication of the confidence associated with the classification of the target file in the malicious class or the non-malicious class, and

wherein the response further comprises the indication of the confidence associated with the classification of the target file in the selected class; and

provide the indication of the confidence, which is associated with the classification of the target file in the selected class, to the user via the user interface.

12 . A method implemented by a computing system, the method comprising:

generating a target vector based at least on an identifier that identifies a target file, content of the target file, and an AI prompt;

determining extents to which historical vectors associated with historical files, which are previously classified by an AI model, correspond to the target vector;

selecting a first historical vector based at least on the extent to which the first historical vector corresponds to the target vector being greater than or equal to an extent threshold;

causing the AI model to perform classification of the target file with regard to maliciousness and to provide an explanation that describes reasoning used by the AI model in performing the classification by providing the AI prompt together with the target file and classification information associated with the first historical vector as inputs to the AI model, the AI prompt requesting the classification of the target file and further requesting the explanation;

receiving a response to the AI prompt from the AI model, the response comprising an indication of a selected class in which the AI model classifies the target file and further comprising the explanation that describes the reasoning used by the AI model in classifying the target file in the selected class, wherein the selected class is a malicious class or a non-malicious class;

storing the explanation, which describes the reasoning used by the AI model in classifying the target file in the selected class; and

based at least on the indication of the selected class in which the AI model classifies the target file, performing a security operation with regard to the target file.

13 . The method of claim 12 , wherein causing the AI model to perform the classification of the target file with regard to maliciousness comprises:

providing the AI prompt together with the target file and contextual information regarding the target file as inputs to the AI model, wherein the contextual information comprises

at least one of the following:

an indication of a signor that signs the target file;

an indication of a publisher that publishes the target file;

a certificate associated with the target file;

at least one of a header byte associated with the target file or a footer byte associated with the target file; or

an indication of behavior associated with the target file.

14 . The method of claim 12 , further comprising:

storing classification information in a store,

wherein the classification information comprises a reference identifier that identifies a reference file, reference content of the reference file, a reference AI prompt requesting classification of the reference file with regard to maliciousness and further requesting a reference explanation that describes reference reasoning used by the AI model in performing the classification of the reference file, and a reference response to the reference AI prompt,

wherein the reference response comprises an indication of a reference class in which the reference file is classified and further comprises the reference explanation that describes the reference reasoning for classifying the reference file in the reference class, wherein the reference class is the malicious class or the non-malicious class; and

in response to storing the classification information in the store, determining that the target file is to be classified among a plurality of classes that comprises the malicious class and the non-malicious class;

wherein causing the AI model to perform the classification of the target file comprises:

causing the AI model to perform the classification of the target file with regard to maliciousness and to provide the explanation that describes the reasoning used by the AI model in performing the classification of the target file by providing the AI prompt together with the target file and the classification information to the AI model.

15 . The method of claim 12 ,

wherein the historical vectors are based at least on identifiers that identify the historical files, contents of the historical files, and AI prompts requesting classifications of the historical files with regard to maliciousness and further requesting explanations that describe reasonings used by the AI model in performing the classifications; and

wherein the classification information comprises a first identifier that identifies a first historical file of the historical files that is associated with the first historical vector, first content of the first historical file, a first AI prompt requesting classification of the first historical file with regard to maliciousness and further requesting a first explanation that describes a first reasoning used by the AI model in performing the classification of the first historical file, and a first response to the first AI prompt from the AI model, and

wherein the first response comprises an indication of a first class in which the AI model classifies the first historical file and further comprises the first explanation that describes the first reasoning used by the AI model in performing the classification of the first historical file, wherein the first class is the malicious class or the non-malicious class.

16 . The method of claim 15 , wherein the historical vectors are further based at least on confidences associated with the classifications of the historical files with regard to maliciousness; and

wherein selecting the first historical vector from the historical vectors comprises:

selecting the first historical vector from the historical vectors based at least on the extent to which the first historical vector corresponds to the target vector being greater than or equal to the extent threshold and further based at least on a first confidence of the confidences that is associated with a first classification of the first historical file with regard to maliciousness being greater than or equal to a confidence threshold.

17 . The method of claim 15 , further comprising:

selecting a second historical vector from the historical vectors based at least on the extent to which the second historical vector corresponds to the target vector being greater than or equal to the extent threshold,

wherein the classification information further comprises a second identifier that identifies a second historical file of the historical files that is associated with the second historical vector, second content of the second historical file, a second AI prompt requesting classification of the second historical file with regard to maliciousness and further requesting a second explanation that describes a second reasoning used by the AI model in performing the classification of the second historical file, and a second response to the second AI prompt from the AI model, and

wherein the second response comprises an indication of a second class in which the AI model classifies the second historical file and further comprises the second explanation that describes the second reasoning used by the AI model in performing the classification of the second historical file, wherein the second class is the malicious class or the non-malicious class.

18 . The method of claim 17 , wherein the first response indicates that the first historical file is classified in the malicious class;

wherein the second response indicates that the second historical file is classified in the non-malicious class; and

wherein providing the AI prompt together with the target file and the classification information to the AI model comprises:

based at least on the first response indicating that the first historical file is classified in the malicious class and further based at least on the second response indicating that the second historical file is classified in the non-malicious class, causing the AI model to process the first historical file prior to processing the second historical file.

19 . The method of claim 17 , wherein the first response indicates that the first historical file is classified in the non-malicious class;

wherein the second response indicates that the second historical file is classified in the malicious class; and

wherein providing the AI prompt together with the target file and the classification information to the AI model comprises:

based at least on the first response indicating that the first historical file is classified in the non-malicious class and further based at least on the second response indicating that the second historical file is classified in the malicious class, causing the AI model to process the first historical file prior to processing the second historical file.

20 . The method of claim 12 ,

wherein the historical vectors are based at least on identifiers that identify the historical files, contents of the historical files, and AI prompts requesting classifications of the historical files with regard to maliciousness and further requesting explanations that describe reasonings used by the AI model in performing the classifications;

wherein selecting the first historical vector comprises:

selecting N first historical vectors, which are associated with historical files that are classified in the malicious class, and N second historical vectors, which are associated with historical files that are classified in the non-malicious class, from the historical vectors to provide 2N sample historical vectors based at least on the extents to which the N first historical vectors and the N second historical vectors correspond to the target vector being greater than or equal to an extent threshold, wherein N is a positive integer; and

selecting P sample historical vectors from the 2N sample historical vectors, wherein P is a positive integer that is less than 2N;

wherein the classification information comprises P identifiers that identify P historical files of the historical files that are associated with the P sample historical vectors, contents of the P historical files, P AI prompts requesting classifications of the P historical files with regard to maliciousness and further requesting explanations that describe reasonings used by the AI model in performing the classifications of the P historical files, and P responses to the P AI prompts from the AI model; and

wherein the P responses comprise indications of selected classes in which the AI model classifies the P historical files and further comprise the explanations that describe the reasonings used by the AI model in performing the classifications of the P historical files, wherein a selected class is the malicious class or the non-malicious class.

21 . The method of claim 12 , wherein causing the AI model to perform the classification of the target file and to provide the explanation comprises:

causing the AI model to perform the classification of the target file with regard to maliciousness, to provide the explanation, and to provide an indication of a confidence associated with the classification of the target file;

wherein the AI prompt further requests the indication of the confidence associated with the classification of the target file;

wherein the response further comprises the indication of the confidence associated with the classification of the target file in the selected class; and

wherein the method further comprises:

providing the indication of the confidence, which is associated with the classification of the target file in the selected class, to a user via a user interface.

22 . A computer program product comprising a computer-readable storage medium having instructions recorded thereon for enabling a processor-based system to perform operations, the operations comprising:

generating a target vector based at least on an identifier that identifies a target file, content of the target file, and an AI prompt;

determining extents to which historical vectors associated with historical files, which are previously classified by an AI model, correspond to the target vector;

selecting a first historical vector based at least on the extent to which the first historical vector corresponds to the target vector being greater than or equal to an extent threshold;

causing the AI model to classify the target file in a malicious class or a non-malicious class and to provide an explanation that describes reasoning used by the AI model in choosing the malicious class or the non-malicious class and to provide an indication of a confidence associated with a classification of the target file in the malicious class or the non-malicious class by providing the AI prompt together with the target file and classification information associated with the first historical vector as inputs to the AI model, the AI prompt requesting that the AI model choose between the malicious class and the non-malicious class and further requesting the explanation that describes the reasoning used by the AI model in choosing the malicious class or the non-malicious class and further requesting the indication of the confidence associated with the classification of the target file in the malicious class or the non-malicious class;

receiving a response to the AI prompt from the AI model, the response comprising an indication of a selected class in which the AI model classifies the target file and further comprising the explanation that describes the reasoning used by the AI model in choosing the selected class and further comprising the indication of the confidence associated with the classification of the target file in the selected class, wherein the selected class is the malicious class or the non-malicious class;

providing the explanation, which describes the reasoning used by the AI model in choosing the selected class, and the indication of the confidence, which is associated with the classification of the target file in the selected class, to a user via a user interface; and

based at least on the indication of the selected class in which the AI model classifies the target file, performing a security operation with regard to the target file.

Assignments (2)

Continuity (1)

References Cited (50)