Input

Time-series x_t

Encoder

(Self-Attention)

Latent

z_t

Decoder

(Self-Attention)

Reconstructed

𝑥̂_t

Discriminator

Disc(x) vs Disc(𝑥̂)

Anomaly Score

a_t = ||x - 𝑥̂||² + (1 - Disc(𝑥̂))

Adversarial Learning

Anomaly Scoring

Windowed Input

Encoded Features

Latent Vector z_t

Reconstructed Sequence

Real x

Fake 𝑥̂