Graph-based classification of elements such as files using a tool such as VirusTotal
A method of determining the level of maliciousness of an element using a directed hypergraph to classify the element based on information aggregated from its locally identified close neighbors, queried in a data base such as VirusTotal (VT). A crawling procedure is used starting from elements needing classification and collecting a set of their neighbors forming neighborhoods. These neighbors are then used to classify the elements. The neural network classifier is able to obtain as input an entire neighborhood. The input includes several feature vectors, one for each element in the neighborhood. In addition, a mapping of interconnections can be provided for each group of elements. Finally, a maliciousness level is provided for the elements in question. For an incriminated file one or more actions can be taken, such as isolating a machine that received the file, killing processes started by the file, removing persistence of the file on the network or affected computer, cleaning infected samples, modifying risk assessment for computer or network, generating a report, collecting additional artifacts, triggering a search for related elements, blocking a user from taking actions and sending information to other IT or security systems. For other element types, some of the above actions are applicable as well. In addition, there are other actions specific to particular element types, e.g. blocking an IP address or a web domain from network access, restricting user authorization, blocking access to an external device, shutting down computers, erasing memory devices, filtering e-mail messages, and many more.
1. A method of classifying an element in question comprising:
generating a hypergraph of nodes, edges and neighboring elements by crawling;
generating recursively sets of neighborhoods of the element in question from the hypergraph;
determining a maliciousness level for the element in question by analyzing the sets of neighborhoods with a trained neural network, wherein each layer of the neural network includes two processing steps:
(a) aggregating information from immediate neighbors of each node and using them to calculate additional features for said node;
(b) feeding all features (both original and aggregated from neighbors) into a fully-connected neural layer, one for each node kind;
generating by a profiler a maliciousness level profile for said element in question;
linking information generated by the profiler relating to the element in question and the maliciousness level profile for the element and, for an incriminated file, responding by taking an action or actions such as isolating a machine that received the file, killing processes started by the file, removing persistence of the file on a network or affected computer, cleaning infected samples, modifying risk assessment for computer or network, generating a report, collecting additional artifacts, triggering a search for related elements, blocking a user from taking actions and sending information to other IT or security systems; or for other element types, any of the above actions as applicable, and in addition, other actions specific to particular element types, e.g. blocking an IP address or a web domain from network access, restricting user authorization, blocking access to an external device, shutting down computers, erasing memory devices, filtering e-mail messages, and many more.
2. The method of claim 1 , further comprising an aggregation function, the aggregation function picking out most extreme evidence from all features over all neighbors.
3. The method of claim 2 , wherein neighbors' features are fed into a fully-connected neural layer, and then a maximum over all neighbors is selected.
4. The method of claim 1 , wherein the data base is VirusTotal (VT).
5. The method of claim 1 , wherein nodes are typed, and nodes of different types may have different features.
6. The method of claim 1 , wherein the hypergraph has only a single source node for each edge.
7. The method of claim 1 , wherein each edge is mapped to a set of target nodes.
8. The method of claim 1 , wherein the neural network is trained by providing a set of seed nodes, crawling for a set of the seed nodes' neighborhoods, building a subgraph from the set of the seed nodes' neighborhoods, and training the neural network by feeding the subgraph to the neural network.
9. A method of determining a maliciousness level of an element in question using a hypergraph of neighbors, the method comprising:
receiving an element of unknown maliciousness level;
generating a hypergraph of neighbor target elements found in the local system, a remote data-base, several remote data-bases or a combination of such systems, the hypergraph containing a set of nodes and a set of edges, wherein the nodes represent the neighbor target elements, and the edges represent connections between the neighbor target elements;
classifying nodes and edges in the hypergraph;
generating by a profiler a maliciousness level profile for the element based on aggregation of nodes and edges in the hypergraph;
linking information generated by the profiler relating to the element in question and the maliciousness level profile for the element and, for an in-criminated file, responding by taking an action or actions such as isolating a machine that received the file, killing processes started by the file, removing persistence of the file on a network or affected computer, cleaning infected samples, modifying risk assessment for computer or network, generating a report, collecting additional artifacts, triggering a search for related elements, blocking a user from taking actions and sending information to other IT or security systems; or for other element types, any of the above actions as applicable, and in addition, other actions specific to particular element types, e.g. blocking an IP address or a web domain from network access, restricting user authorization, blocking access to an external device, shutting down computers, erasing memory devices, filtering e-mail messages, and many more.
10. The method of claim 9 , wherein there are from 1 to V target nodes associated with each edge, where V is the number of nodes in the hypergraph.
11. The method of claim 9 , wherein the hypergraph has only a single source node for each edge.
12. The method of claim 9 , wherein the hypergraph contains at least some edges having multiple source nodes.
13. The method of claim 9 , wherein the input data is a hybrid hypergraph comprising of combined data collected locally and data from one or more external data bases (such as VT).
14. The method of claim 9 , wherein the element in question is of known maliciousness level.
15. The method of claim 9 , wherein nodes are typed, and nodes of different types may have different features.
16. The method of claim 9 , wherein sets of neighborhoods of target nodes are generated recursively based on nodes satisfying particular features.
17. The method of claim 9 , wherein the hypergraph is generated by crawling.
18. The method of claim 9 , wherein a set of seed nodes in the hypergraph is classified by a trained neural network.
19. The method of claim 9 , wherein an element may be classified based on its connections to neighbors without a classification of said neighbors.
20. The method of claim 9 , wherein the searching is in VirusTotal (VT).