Granted Patent B2

US 12,386,978 · App. 18/079,665 · Granted Aug 12, 2025

Inventors: Tom Ganz (Karlsruhe, DE); Martin Haerterich (Wiesloch, DE); Philipp Rall (Darmstadt, DE)

Assignee: SAP SE

G06F21/577G06F11/3698G06F21/563G06F2221/033

Abstract

Claims (77)

1. A system comprising:

a memory that stores instructions; and

one or more processors configured by the instructions to perform operations comprising:

training a machine-learning model to detect whether source code contains a vulnerability;

determining, using the trained machine-learning model, whether the vulnerability exists in a source code;

identifying, using a first explanation method, one or more first target locations corresponding to the vulnerability within the source code;

identifying, using a second explanation method, one or more second target locations corresponding to the vulnerability within the source code;

using guided fuzzing, providing input to the source code targeting the first target locations to empirically measure whether the vulnerability exists at the first target locations by running the source code in a debugger with breakpoints set at the one or more first target locations;

using guided fuzzing, providing input to the source code targeting the second target locations to empirically measure whether the vulnerability exists at the second target locations by running the source code in the debugger with breakpoints set at the one or more second target locations;

based on results from the guided fuzzing, selecting between the first target locations and the second target locations based on mean breakpoint hits when running the source code in the debugger; and

indicating, on a user interface, the selected target locations.

2. The system of claim 1 , wherein the identifying of the one or more first target locations comprises:

generating, by the one or more processors, based on the source code and the determination of whether the vulnerability exists in the source code, a map that attributes numerical relevance scores to locations in the source code;

sorting the locations of the map by relevance; and

converting the locations to line numbers in the source code.

3. The system of claim 1 , wherein the empirical measuring of whether the vulnerability exists at the first target locations or the second target locations is further based on a mean crash distance between the one or more target locations and locations where the source code crashes.

4. The system of claim 1 , wherein the identifying of the one or more first target locations in the source code that may cause the vulnerability comprises:

converting the source code to a graph representation;

providing the graph representation to the first explanation method;

receiving one or more graph locations from the first explanation method; and

converting the one or more graph locations to one or more line numbers in the source code.

5. The system of claim 1 , wherein the operations further comprise:

modifying the explanation method based on the source code and the empirical measuring of whether the vulnerability exists at the first target locations or the second target locations.

6. The system of claim 1 , wherein the determining, using the trained machine-learning model, whether the vulnerability exists in the source code comprises:

providing the source code as input to the trained machine-learning model; and

receiving, as output from the trained machine-learning model, a vector comprising a plurality of elements, each element indicating a probability of a corresponding vulnerability existing in the input.

7. The system of claim 1 , wherein the providing input to the source code targeting the first target locations to empirically measure whether the vulnerability exists at the first target locations comprises executing the source code multiple times with different inputs generated by a fuzzing module.

8. A method comprising:

training a machine-learning model to detect whether source code contains a vulnerability;

determining, by one or more processors using the trained machine-learning model, whether the vulnerability exists in a source code;

identifying, using a first explanation method, one or more first target locations corresponding to the vulnerability within the source code;

identifying, using a second explanation method, one or more second target locations corresponding to the vulnerability within the source code;

using guided fuzzing, providing input to the source code targeting the first target locations to empirically measure whether the vulnerability exists at the first target locations by running the source code in a debugger with breakpoints set at the one or more first target locations;

using guided fuzzing, providing input to the source code targeting the second target locations to empirically measure whether the vulnerability exists at the second target locations by running the source code in the debugger with breakpoints set at the one or more second target locations;

based on results from the guided fuzzing, selecting between the first target locations and the second target locations based on mean breakpoint hits when running the source code in the debugger; and

indicating, on a user interface, the selected target locations.

9. The method of claim 8 , wherein the identifying of the one or more first target locations comprises:

generating, by the one or more processors, based on the source code and the determination of whether the vulnerability exists in the source code, a map that attributes numerical relevance scores to locations in the source code;

sorting the locations of the map by relevance; and

converting the locations to line numbers in the source code.

10. The method of claim 8 , wherein the empirical measuring of whether the vulnerability exists at the first target locations or the second target locations is further based on a mean crash distance between the one or more target locations and locations where the source code crashes.

11. The method of claim 8 , wherein the identifying of the one or more first target locations in the source code that may cause the vulnerability comprises:

converting the source code to a graph representation;

providing the graph representation to the first explanation method;

receiving one or more graph locations from the first explanation method; and

converting the one or more graph locations to one or more line numbers in the source code.

12. The method of claim 8 , further comprising:

modifying the explanation method based on the source code and the empirical measuring of whether the vulnerability exists at the first target locations or the second target locations.

13. The method of claim 8 , wherein the determining, using the trained machine-learning model, whether the vulnerability exists in the source code comprises:

providing the source code as input to the trained machine-learning model; and

receiving, as output from the trained machine-learning model, a vector comprising a plurality of elements, each element indicating a probability of a corresponding vulnerability existing in the input.

14. The method of claim 8 , wherein the providing input to the source code targeting the first target locations to empirically measure whether the vulnerability exists at the first target locations comprises executing the source code multiple times with different inputs generated by a fuzzing module.

15. A non-transitory computer-readable medium that stores instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

training a machine-learning model to detect whether source code contains a vulnerability;

determining, using the trained machine-learning model, whether the vulnerability exists in a source code;

identifying, using a first explanation method, one or more first target locations corresponding to the vulnerability within the source code;

identifying, using a second explanation method, one or more second target locations corresponding to the vulnerability within the source code;

using guided fuzzing, providing input to the source code targeting the first target locations to empirically measure whether the vulnerability exists at the first target locations by running the source code in a debugger with breakpoints set at the one or more first target locations;

using guided fuzzing, providing input to the source code targeting the second target locations to empirically measure whether the vulnerability exists at the second target locations by running the source code in the debugger with breakpoints set at the one or more second target locations;

based on results from the guided fuzzing, selecting between the first target locations and the second target locations based on mean breakpoint hits when running the source code in the debugger; and

indicating, on a user interface, the selected target locations.

16. The non-transitory computer-readable medium of claim 15 , wherein the identifying of the one or more first target locations comprises:

generating, by the one or more processors, based on the source code and the determination of whether the vulnerability exists in the source code, a map that attributes numerical relevance scores to locations in the source code;

sorting the locations of the map by relevance; and

converting the locations to line numbers in the source code.

17. The non-transitory computer-readable medium of claim 15 , wherein the empirical measuring of whether the vulnerability exists at the first target locations or the second target locations is further based on a mean crash distance between the one or more target locations and locations where the source code crashes.

18. The non-transitory computer-readable medium of claim 15 , wherein the operations further comprise:

modifying the explanation method based on the source code and the empirical measuring of whether the vulnerability exists at the first target locations or the second target locations.

19. The non-transitory computer-readable medium of claim 15 , wherein the identifying of the one or more first target locations in the source code that may cause the vulnerability comprises:

converting the source code to a graph representation;

providing the graph representation to the first explanation method;

receiving one or more graph locations from the first explanation method; and

converting the one or more graph locations to one or more line numbers in the source code.

20. The non-transitory computer-readable medium of claim 15 , wherein the determining, using the trained machine-learning model, whether the vulnerability exists in the source code comprises:

providing the source code as input to the trained machine-learning model; and

receiving, as output from the trained machine-learning model, a vector comprising a plurality of elements, each element indicating a probability of a corresponding vulnerability existing in the input.

21. The non-transitory computer-readable medium of claim 15 , wherein the providing input to the source code targeting the first target locations to empirically measure whether the vulnerability exists at the first target locations comprises executing the source code multiple times with different inputs generated by a fuzzing module.

Assignments (2)

Continuity (2)

References Cited (60)