Adversarial Anomaly Detection

Unsupervised Adversarial Anomaly Detection

Task: Anomaly detection in multivariate time series variable.

• About 150 features in 1 input sample per 1 sec.
• t1: [f1, f2, ... , f150]
• t2: [f1, f2, ... , f150]
• ...
• tn: [f1, f2, ... , f150]

SEADNet

• Sequence Embedding, Anomaly Detection Network

Note) I assumed you have 2 GPUs. If not, change gpu setting code.

Stacked LSTM + Deep Auto Encoder + Adversarial Discriminator + Latent Reencoder

Input

• Muti-variate time series (Feature dimension: 150, 1 Event per sec)

Output

• Anomaly score

Stacked LSTM

• Many to One
• Create sequence latent for sequence anomaly
• Sequence length: 20

Deep Auto Encoder

• Create local anomaly latent

Decoder

• Concat sequence latent and local latent
• Reconstruct using concatenated latent variable

Test

• Normal: mean 0.0, stdev 1.0, 100,000 samples, sigmoid standardazation for harder discrimination
• Anomaly: mean 1.0, stddev 1.0, 100 samples, sigmoid standardazation for harder discrimination
• Training only normal samples
• Discriminate anomaly samples at test time

STADNet

• Spatio-Temporal, Anomaly Detection Network

Note) I assumed you have 2 GPUs. If not, change gpu setting code.

1D/2D Convolution + Adversarial Discriminator

Input

• Muti-variate time series (Feature dimension: 150, Window size: 20)
• Reshape to [B, 150, 20, 1]

Output

• Anomaly score

Deep Auto Encoder

• Create local anomaly latent

Decoder

• Concat sequence latent and local latent
• Reconstruct using concatenated latent variable

Test

• Normal: mean 0.0, stdev 1.0, 100,000 samples, sigmoid standardazation for harder discrimination
• Anomaly: mean 1.0, stddev 1.0, 100 samples, sigmoid standardazation for harder discrimination
• Training only normal samples
• Discriminate anomaly samples at test time

Always welcome a nice idea: Code or Text what ever!