Structure-aware neural networks for malware detection
Provided is a malware detection system that provides structure-aware neural networks for performing malware detection. In particular, rather than treat the entire computer file as one large input to a deep neural network, the malware detection system can break the file up based on the internal file structure. Each portion of the computer file can then be processed using individual neural networks and the outputs of these networks can be combined and similarly processed. In this way the overall system can evaluate the file with knowledge of the structure of the file, enabling the malware detection to have a higher-order understanding of the interoperation of different portions of the computer file.
1 . A computer-implemented method to perform malware detection via hierarchically-organized machine learning models, the method comprising:
obtaining, by a computing system comprising one or more computing devices, a computer file to be analyzed for malware detection, wherein the computer file has a structure;
parsing, by the computing system, the computer file into a plurality of file portions based on the structure of the computer file;
respectively processing, by the computing system, the plurality of file portions with a plurality of machine-learned feature extractor models to respectively generate a plurality of feature embeddings respectively for the plurality of file portions, such that each file portion is processed by one of the machine-learned feature extractor models to generate one of the feature embeddings, and wherein each feature embedding comprises a numerical vector expressed within a latent space;
combining, by the computing system, the plurality of feature embeddings within the latent space; and
processing, by the computing system, the combination of the plurality of feature embeddings with a machine-learned prediction model to generate, as an output of the machine-learned prediction model, a model prediction for the computer file, wherein the model prediction is indicative of a malware classification for the computer file.
2 . The computer-implemented method of claim 1 , wherein:
parsing, by the computing system, the computer file into the plurality of file portions based on the structure of the computer file comprises assigning, by the computing system, a respective portion class to each of the plurality of file portions; and
the respective machine-learned feature extractor model used to process each file portion comprises a class-specific model trained specifically for the portion class assigned to the file portion.
3 . The computer-implemented method of claim 1 , wherein parsing, by the computing system, the computer file into the plurality of file portions based on the structure of the computer file comprises applying, by the computing system, a rule-based parsing engine that applies a plurality of user-defined parsing rules.
4 . The computer-implemented method of claim 1 , wherein each of the plurality of machine-learned feature extractor models comprises a convolutional neural network, and wherein the machine-learned prediction model comprises a transformer neural network.
5 . The computer-implemented method of claim 1 , wherein, for at least a first file portion of the plurality of file portions:
the method further comprises parsing, by the computing system, the first file portion into a plurality of file sub-portions; and
processing, by the computing system, the first file portion with the respective machine-learned feature extractor model comprises:
respectively processing, by the computing system, the plurality of file sub-portions with a plurality of machine-learned feature extractor sub-models to respectively generate a plurality of intermediate feature embeddings respectively for the plurality of file sub-portions; and
processing, by the computing system, the plurality of intermediate feature embeddings with a feature aggregation model to generate the feature embedding for the first file portion.
6 . The computer-implemented method of claim 1 , wherein the model prediction comprises a binary or logistic malware classification for the computer file.
7 . The computer-implemented method of claim 1 , wherein the model prediction comprises a multiclass malware family classification for the computer file.
8 . The computer-implemented method of claim 1 , wherein the model prediction comprises a functionality classification or an attribution classification for the computer file.
9 . The computer-implemented method of claim 1 , further comprising:
determining, by the computing system, that the computer file comprises malware based on the model prediction; and
in response to determining that the computer file comprises malware: generating, by the computing system, a malware alert or performing, by the computing system, a malware response action.
10 . One or more non-transitory computer-readable media that collectively store instructions that, when executed by one or more processors of a computing system, cause the computing system to perform operations, the operations comprising:
obtaining, by the computing system, a computer file, wherein the computer file has a structure;
parsing, by the computing system, the computer file into a plurality of file portions based on the structure of the computer file;
respectively processing, by the computing system, the plurality of file portions with a plurality of machine-learned feature extractor models to respectively generate a plurality of feature embeddings respectively for the plurality of file portions, such that each file portion is processed by one of the machine-learned feature extractor models to generate one of the feature embeddings, and wherein each feature embedding comprises a numerical vector expressed within a latent space;
combining, by the computing system, the plurality of feature embeddings within the latent space; and
processing, by the computing system, the combination of the plurality of feature embeddings with a machine-learned prediction model to generate, as an output of the machine-learned prediction model, an output representation for the computer file.
11 . The one or more non-transitory computer-readable media of claim 10 , wherein:
parsing, by the computing system, the computer file into the plurality of file portions based on the structure of the computer file comprises assigning, by the computing system, a respective portion class to each of the plurality of file portions; and
the respective machine-learned feature extractor model used to process each file portion comprises a class-specific model trained specifically for the portion class assigned to the file portion.
12 . The one or more non-transitory computer-readable media of claim 10 , wherein parsing, by the computing system, the computer file into the plurality of file portions based on the structure of the computer file comprises applying, by the computing system, a rule-based parsing engine that applies a plurality of user-defined parsing rules.
13 . The one or more non-transitory computer-readable media of claim 10 , wherein each of the plurality of machine-learned feature extractor models comprises a convolutional neural network and the machine-learned prediction model comprises a transformer neural network.
14 . The one or more non-transitory computer-readable media of claim 10 , wherein, for at least a first file portion of the plurality of file portions:
the operations further comprise parsing, by the computing system, the first file portion into a plurality of file sub-portions; and
processing, by the computing system, the first file portion with the respective machine-learned feature extractor model comprises:
respectively processing, by the computing system, the plurality of file sub-portions with a plurality of machine-learned feature extractor sub-models to respectively generate a plurality of intermediate feature embeddings respectively for the plurality of file sub-portions; and
processing, by the computing system, the plurality of intermediate feature embeddings with a feature aggregation model to generate the feature embedding for the first file portion.
15 . The one or more non-transitory computer-readable media of claim 10 , further comprising:
using, by the computing system, the output representation for the computer file to perform a downstream task for the computer file, wherein the downstream task comprises a similarity search task or an anomaly detection task.
16 . A computing system to train a set of hierarchically-organized machine learning models for malware detection, the computer system comprising one or more processors and one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:
obtaining, by the computing system, a computer file, wherein the computer file has a structure, and wherein a ground truth label is associated with the computer file;
parsing, by the computing system, the computer file into a plurality of file portions based on the structure of the computer file;
respectively processing, by the computing system, the plurality of file portions with a plurality of machine-learned feature extractor models to respectively generate a plurality of feature embeddings respectively for the plurality of file portions, such that each file portion is processed by one of the machine-learned feature extractor models to generate one of the feature embeddings, and wherein each feature embedding comprises a numerical vector expressed within a latent space;
combining, by the computing system, the plurality of feature embeddings within the latent space; and
processing, by the computing system, the combination of the plurality of feature embeddings with a machine-learned prediction model to generate, as an output of the machine-learned prediction model, a model prediction for the computer file, wherein the model prediction is indicative of a malware classification for the computer file;
evaluating, by the computing system, a loss function that generates a loss value based on a comparison of the model prediction for the computer file with the ground truth label for computer file; and
modifying, by the computing system, one or more parameter values of the machine-learned prediction model and the plurality of machine-learned feature extractor models based on the loss function.
17 . The computing system of claim 16 , wherein:
parsing, by the computing system, the computer file into the plurality of file portions based on the structure of the computer file comprises assigning, by the computing system, a respective portion class to each of the plurality of file portions; and
the respective machine-learned feature extractor model used to process each file portion comprises a class-specific model trained specifically for the portion class assigned to the file portion.
18 . The computing system of claim 16 , wherein parsing, by the computing system, the computer file into the plurality of file portions based on the structure of the computer file comprises applying, by the computing system, a rule-based parsing engine that applies a plurality of user-defined parsing rules.
19 . The computing system of claim 16 , wherein each of the plurality of machine-learned feature extractor models comprises a convolutional neural network and the machine-learned prediction model comprises a transformer neural network.
20 . The computing system of claim 16 , wherein, for at least a first file portion of the plurality of file portions:
the operations further comprise parsing, by the computing system, the first file portion into a plurality of file sub-portions; and
processing, by the computing system, the first file portion with the respective machine-learned feature extractor model comprises:
respectively processing, by the computing system, the plurality of file sub-portions with a plurality of machine-learned feature extractor sub-models to respectively generate a plurality of intermediate feature embeddings respectively for the plurality of file sub-portions; and
processing, by the computing system, the plurality of intermediate feature embeddings with a feature aggregation model to generate the feature embedding for the first file portion.