Sensitive data leakage protection
View Patent ↗Disclosed herein are system, method, and computer program product embodiments for increasing data security by using generative adversarial networks (GAN) and transformer models to detect sensitive data leakage. A transformer model may receive a message via a network. The transformer model may then apply a GAN model to determine whether the message contains potentially sensitive data requiring further inspection. If the message contains potentially sensitive data, the transformer model may determine sensitive data types may be included in the message and the likelihood that sensitive data is present. Based on the transformer model's determination, a blocking policy may be applied to prevent sensitive data leakage.
1 . A computer implemented method for sensitive data leakage protection, the method comprising:
receiving a message having a destination residing outside of a secure network;
determining that the message contains a potentially sensitive data component by applying a generative adversarial network (GAN) model to the message;
transforming the message into a message vector, wherein the message vector is based on content of the message;
determining, by a transformer model, that the potentially sensitive data component includes sensitive data by:
applying the message vector to the transformer model;
comparing the message vector to one or more vectors stored in a sensitive data database, wherein each of the one or more vectors corresponds to a type of sensitive data; and
in response to the comparison, identifying a type of sensitive data;
applying a blocking policy to the message based on the type of sensitive data;
generating a sensitive data report including one or more sensitive data types respectively corresponding to each of the one or more vectors and a similarity value of each of the one or more vectors to the message vector; and
generating a graphical user interface that includes the sensitive data report.
2 . The computer implemented method of claim 1 , wherein the GAN model comprises a message generator and a message discriminator, wherein the method further comprises:
generating, by the message generator, a false positive message;
determining, by the message discriminator, whether the false positive message was created by the message generator;
re-training the message generator based on whether the message discriminator correctly determined that the false positive message was created by the message generator; and
re-training the message discriminator based on whether the message discriminator correctly determined that the false positive message was created by the message generator.
3 . The computer implemented method of claim 1 , wherein comparing the message vector to the one or more vectors stored in the sensitive data database further comprises:
determining a similarity value between the message vector and each of the one or more vectors stored in the sensitive data database.
4 . The computer implemented method of claim 3 , wherein each type of sensitive data has a corresponding similarity threshold, and wherein applying the blocking policy further comprises:
determining that a similarity value between the message vector and a vector stored in the sensitive data database is greater than the similarity threshold corresponding to the vector.
5 . The computer implemented method of claim 3 , wherein the similarity value is determined by applying a cosine similarity or nearest neighbor search.
6 . The computer implemented method of claim 1 , wherein the sensitive data comprises a credit card number, a social security number, a username, a password, a passport number, a driver's license number, or an account number.
7 . The computer implemented method of claim 1 , wherein applying the blocking policy comprises discarding the message, storing the message in a database, or removing the sensitive data component from the message.
8 . A system, comprising:
a memory; and
at least one processor coupled to the memory and configured to:
receive a message having a destination residing outside of a secure network;
determine that the message contains a potentially sensitive data component by applying a generative adversarial network (GAN) model to the message;
transform, by a transformer model, the message into a message vector, wherein the message vector is based on content of the message;
determine, by the transformer model, that the potentially sensitive data component is sensitive by:
applying the message vector to the transformer model;
comparing the message vector to one or more vectors stored in a sensitive data database, wherein each of the one or more vectors corresponds to a type of sensitive data; and
in response to the comparison, identifying a type of sensitive data;
apply a blocking policy to the message based on the type of sensitive data;
generate a sensitive data report including one or more sensitive data types respectively corresponding to each of the one or more vectors and a similarity value of each of the one or more vectors to the message vector; and
generate a graphical user interface that includes the sensitive data report.
9 . The system of claim 8 , wherein the GAN model comprises a message generator and a message discriminator, wherein the at least one processor is further configured to:
generate, by the message generator, a false positive message;
determine, by the message discriminator, whether the false positive message was created by the message generator;
re-train the message generator based on whether the message discriminator correctly determined that the false positive message was created by the message generator; and
re-train the message discriminator based on whether the message discriminator correctly determined that the false positive message was created by the message generator.
10 . The system of claim 8 , wherein to compare the message vector to the one or more vectors stored in the sensitive data database, the at least one processor is further configured to:
determine a similarity value between the message vector and each of the one or more vectors stored in the sensitive data database.
11 . The system of claim 10 , wherein each type of sensitive data has a corresponding similarity threshold, and wherein applying the blocking policy further comprises:
determining that a similarity value between the message vector and a vector stored in the sensitive data database is greater than the similarity threshold corresponding to the vector.
12 . The system of claim 10 , wherein the similarity value is determined by applying a cosine similarity or nearest neighbor search.
13 . The system of claim 8 , wherein the sensitive data comprises a credit card number, a social security number, a username, a password, a passport number, a driver's license number, or an account number.
14 . The system of claim 9 , wherein applying the blocking policy comprises discarding the message, storing the message in a database, or removing the sensitive data component from the message.
15 . A non-transitory computer-readable device having instructions stored thereon that, when executed by at least one computing device, cause the at least one computing device to perform operations comprising:
receiving a message having a destination residing outside of a secure network;
determining that the message contains a potentially sensitive data component by applying a generative adversarial network (GAN) model to the message;
transforming, by a transformer model, the message into a message vector, wherein the message vector is based on content of the message;
determining, by the transformer model, that the potentially sensitive data component includes sensitive data by:
applying the message vector to the transformer model;
comparing the message vector to one or more vectors stored in a sensitive data database, wherein each of the one or more vectors corresponds to a type of sensitive data; and
in response to the comparison, identifying a type of sensitive data;
applying a blocking policy to the message based on the type of sensitive data;
generating a sensitive data report including one or more sensitive data types respectively corresponding to each of the one or more vectors and a similarity value of each of the one or more vectors to the message vector; and
generating a graphical user interface that includes the sensitive data report.
16 . The non-transitory computer-readable device of claim 15 , wherein the GAN model comprises a message generator and a message discriminator, wherein the operations further comprise:
generating, by the message generator, a false positive message;
determining, by the message discriminator, whether the false positive message was created by the message generator;
re-training the message generator based on whether the message discriminator correctly determined that the false positive message was created by the message generator; and
re-training the message discriminator based on whether the message discriminator correctly determined that the false positive message was created by the message generator.
17 . The non-transitory computer-readable device of claim 15 , wherein comparing the message vector to the one or more vectors stored in the sensitive data database further comprises:
determining a similarity value between the message vector and each of the one or more vectors stored in the sensitive data database.
18 . The non-transitory computer-readable device of claim 17 , wherein each type of sensitive data has a corresponding similarity threshold, and wherein applying the blocking policy further comprises:
determining that a similarity value between the message vector and a vector stored in the sensitive data database is greater than the similarity threshold corresponding to the vector, wherein the similarity value is determined by applying a cosine similarity or nearest neighbor search.