Explainable neural network for anomaly detection
Disclosed are a process, system and medium for explaining an anomaly detected in an authentication request by a classifier, including. obtaining the anomaly detection model, with the anomaly detection model having been trained to detect anomalous authentication requests. Also disclosed is obtaining a multi-layer perceptron (MLP) model trained to provide MLP results similar to anomaly detection results when the same features are provided to both the anomaly detection model and the MLP model. During the MLP model training, the anomaly detection model results serve as ground truth; also, detecting, by the anomaly detection model, that the authentication request is anomalous and providing the authentication request features to an explanation tool using the MLP. Based on the detecting, also disclosed is obtaining, from the explanation tool, an explanation comprising at least one identification of the authentication request feature that contributed most to the detection of anomaly.
1 . A method of explaining an anomaly detected in an authentication request by a classifier, the method comprising:
obtaining an unsupervised encoder-decoder model, the encoder-decoder model having been trained to detect authentication requests;
obtaining a multi-layer perceptron (MLP) model trained to provide MLP results similar to encoder-decoder results when the same features are provided to both the unsupervised encoder-decoder model and the MLP model, wherein the encoder-decoder results served as ground truth during the MLP model training, the MLP model is trained by providing, as input, the input to the encoder-decoder model, and targeting, as output, a reconstruction loss error of the encoder-decoder model;
embedding authentication request features from an authentication request into an embedding space to generate embedding features;
providing the embedding features to the encoder-decoder model;
detecting, by the encoder-decoder model, that the authentication request is anomalous, the anomaly is reflected by a risk score from the encoder-decoder model, and an explainability condition is satisfied when the risk score exceeds a threshold;
responsive to the detecting the authentication request as anomalous, providing the authentication request features to an explanation tool using the MLP model; and
obtaining, from the explanation tool, an explanation comprising at least an identification of at least one of the authentication request features that contributed most to the detecting the authentication request as anomalous.
2 . The method of claim 1 , wherein the explanation tool is integrated gradients (IG).
3 . The method of claim 1 , wherein the encoder-decoder model is a variational autoencoder (VAE).
4 . The method of claim 1 , wherein each authentication request feature in the explanation is paired with a score and the score reflects a degree of contribution to the detecting by the encoder-decoder model.
5 . The method of claim 4 , wherein a magnitude of the score reflects the degree of contribution.
6 . A non-transitory computer readable storage medium impressed with computer program instructions to explain an anomaly detected in an authentication request by a classifier, the instructions, when executed on a processor, implement a method comprising:
obtaining an unsupervised encoder-decoder model, the encoder-decoder model having been trained to detect anomalous authentication requests;
obtaining a multi-layer perceptron (MLP) model trained to provide MLP results similar to encoder-decoder results when the same features are provided to both the unsupervised encoder-decoder model and the MLP model, wherein the encoder-decoder results served as ground truth during the MLP model training, the MLP model is trained by providing, as input, the input to the encoder-decoder model, and targeting, as output, a reconstruction loss error of the encoder-decoder model;
embedding authentication request features from an authentication request into an embedding space to generate embedding features;
providing the embedding features to the encoder-decoder model;
detecting, by the encoder-decoder model, that the authentication request is anomalous, the anomaly is reflected by a risk score from the encoder-decoder model, and an explainability condition is satisfied when the risk score exceeds a threshold;
responsive to the detecting the authentication request as anomalous, providing the authentication request features to an explanation tool using the MLP model; and
obtaining, from the explanation tool, an explanation comprising at least an identification of at least one of the authentication request features that contributed most to the detecting the authentication request as anomalous.
7 . The non-transitory computer readable storage medium of claim 6 , wherein the explanation tool is integrated gradients (IG).
8 . The non-transitory computer readable storage medium of claim 6 , wherein the encoder-decoder model is a variational autoencoder (VAE).
9 . The non-transitory computer readable storage medium of claim 6 , wherein each authentication request feature in the explanation is paired with a score and the score reflects a degree of contribution to the detecting by the encoder-decoder model.
10 . The non-transitory computer readable storage medium of claim 9 , wherein a magnitude of the score reflects the degree of contribution.
11 . A system including one or more processors coupled to memory, the memory loaded with computer instructions to explain an anomaly detected in an authentication request by a classifier, the instructions, when executed on the one or more processors, implement actions comprising:
obtaining an unsupervised encoder-decoder model, the encoder-decoder model having been trained to detect anomalous authentication requests;
obtaining a multi-layer perceptron (MLP) model trained to provide MLP results similar to encoder-decoder results when the same features are provided to both the unsupervised encoder-decoder model and the MLP model, wherein the encoder-decoder results served as ground truth during the MLP model training, the MLP model is trained by providing, as input, the input to the encoder-decoder model, and targeting, as output, a reconstruction loss error of the encoder-decoder model;
embedding authentication request features from an authentication request into an embedding space to generate embedding features;
providing the embedding features to the encoder-decoder model;
detecting, by the encoder-decoder model, that the authentication request is anomalous, the anomaly is reflected by a risk score from the encoder-decoder model, and an explainability condition is satisfied when the risk score exceeds a threshold;
responsive to the detecting the authentication request as anomalous, providing the authentication request features to an explanation tool using the MLP model; and
obtaining, from the explanation tool, an explanation comprising at least an identification of at least one of the authentication request features that contributed most to the detecting the authentication request as anomalous.
12 . The system of claim 11 , wherein the explanation tool is integrated gradients (IG).
13 . The system of claim 11 , wherein the encoder-decoder model is a variational autoencoder (VAE).
14 . The system of claim 11 , wherein each authentication request feature in the explanation is paired with a score and the score reflects a degree of contribution to the detecting by the encoder-decoder model.