Phishing URL detection using transformers
The technology described herein can identify phishing URLs using transformers. The technology tokenizes useful features from the subject URL. The useful features can include the text of the URL and other data associated with the URL, such as certificate data for the subject URL, a referrer URL, an IP address, etc. The technology may build a joint Byte Pair Encoding for the features. The token encoding may be processed through a transformer, resulting in a transformer output. The transformer output, which may be described as a token embedding, may be input to a classifier to determine whether the URL is a phishing URL. Additional or improved URL training data may be generated by permuting token order, by simulating a homoglyph attack, and by simulating a compound word attack.
1. One or more computer storage media having computer-executable instructions embodied thereon that, when executed, by one or more processors, causes the one or more processors to perform a method comprising:
generating a set of Uniform Resource Locator (URL) tokens based on a URL;
generating a set of metadata tokens based on metadata associated with the URL;
generating a set of feature tokens based on the set of URL tokens, the set of metadata tokens, and a set of separator tokens by at least concatenating the set of URL tokens and the set of metadata tokens into the set of feature tokens including a first separator token from the set of separator tokens between a first metadata token of the set of metadata tokens and a second metadata token, wherein the first separator token indicates a type of metadata associated with the second metadata token;
providing the set of feature tokens as a single input vector to a transformer model;
obtaining an output of the transformer model including an embedding vector;
determining a decision statistic based on the embedding vector; and
as a result of the decision statistic indicating the URL is malicious, causing a remedial action to be performed, where the remedial action prevents a computing device from accessing the URL.
2. The media of claim 1 , wherein the set of feature tokens comprises a token string for a subject of the URL, a token string for source context data associated with the URL, a token string for a referring URL, a token string for destination context data, associated with the URL a token string for certificate context associated with the URL, and a token string for IP address associated with the URL separated using the separator tokens of the set of separator tokens.
3. The media of claim 1 , the first metadata token and the second metadata token are associated with different types of metadata.
4. The media of claim 1 , wherein the transformer model is trained using adversarial training to augment a training dataset.
5. The media of claim 1 , wherein the remedial action further comprises preventing the computing device from accessing the URL by at least redirecting the computing device to a second URL.
6. The media of claim 1 , wherein determining the decision statistic based on the embedding vector further comprises adding weight values to the output of the transformer model.
7. The media of claim 1 , wherein providing the set of feature tokens as the input to the transformer model further comprises generating a set of token encodings based on the set of feature tokens, the set of token encodings including word tokens and corresponding positions.
8. A method for comprising:
generating a set of feature tokens based on a set of Uniform Resource Locator (URL) tokens based on a URL, a set of metadata tokens based on metadata associated with the URL, and a set of separator tokens, wherein the set of feature tokens includes a first separator token between a first URL token and a first metadata token, where the first separator token indicates a type associated with the first metadata token;
providing the set of feature tokens as a single input vector to a transformer model;
obtaining an output of the transformer model; and
based on the output causing a remedial action to be performed, wherein the remedial action mitigates a risk associated with the URL.
9. The method of claim 8 , wherein generating the set of feature tokens further comprises including separator tokens of the set of separator tokens between URL tokens of the set of URL tokens and metadata tokens of the set of metadata tokens.
10. The method of claim 9 , wherein generating the set of feature tokens further comprises including the separator tokens of the set of separator tokens between the metadata tokens of the set of metadata tokens.
11. The method of claim 10 , wherein each separator token of the set of separator tokens indicates a different type of metadata associated within the metadata tokens that follows each separator token.
12. The method of claim 11 , wherein the first separator token indicates a referral URL and a second separator token of the set of separator tokens indicates a certificate authority.
13. The method of claim 8 , wherein the transformer model is a bidirectional encoder representation trained with URL training data.
14. The method of claim 8 , wherein providing the set of feature tokens as the single input vector to the transformer model further comprises generating a token encoding using a Byte Pair Encoding (BPE) model.
15. The method of claim 8 , comprising preventing a computing device from accessing the URL.
16. A system comprising:
one or more hardware processors; and
one or more computer-readable media having executable instructions embodied thereon, which, when executed by the one or more hardware processors, cause the one or more hardware processors to execute actions comprising:
generating a feature token encoding based on a set of Uniform Resource Locator (URL) tokens extracted from a URL, a set of metadata tokens extracted from metadata associated with the URL, and including a set of separator tokens between tokens of the feature token encoding, where separator tokens of the set of separator tokens indicate types of metadata that follow the separator tokens;
causing a transformer model to perform inferencing by at least providing as a single input vector to the transformer model the feature token encoding;
obtaining an output of the transformer model indicating a score associated with the URL; and
based on the score indicating the URL is malicious, causing a remedial action to be performed, where the remedial action limits access to the URL.
17. The system of claim 16 , wherein the actions further comprise determining the score by at least applying a set of weights to values included in an embedding vector including the output.
18. The system of claim 17 , wherein the transformer model is a bidirectional encoder representation.
19. The system of claim 16 , wherein the score is determined with a machine classifier comprising a single fully connected layer.
20. The system of claim 16 , wherein the feature token encoding includes at least one of: a token representing a subject of the URL, a token representing source metadata associated with the URL, a token representing a referring URL, a token representing destination metadata, a token representing a certificate authority, and a token representing an IP address associated with the URL.