Reproduce “Efficient GAN-Based Anomaly Detection” on OpenResearch

About this paper

Houssam Zenati, Chuan Sheng Foo, Bruno Lecouat, Gaurav Manek, Vijay Ramaseshan Chandrasekhar

The authors develop an encoder-based GAN for anomaly detection in images and network activity. Jointly training an encoder and generator enables test inputs to be reconstructed and scored without per-example latent-code optimization. On KDD99, its cross-entropy variant reports F1 of 0.9372, nearly tied with DAGMM’s 0.9369; inference takes 2.7 ms versus AnoGAN’s 2,578 ms on reported hardware.

Key insights

• Integrating an encoder directly into the GAN training (as in BiGAN/ALI) provides a direct, single-pass mapping from data to latent space, dramatically accelerating anomaly detection inference.
• Combining reconstruction error with a discriminator-based loss (which assesses if a data-latent pair is "real") yields a more robust anomaly score than either component alone.
• The optimal variant of the discriminator-based loss (cross-entropy vs. feature-matching) can be data-dependent, with feature-matching performing better on image data (MNIST) and cross-entropy performing slightly better on tabular data (KDD99).

Summary by alphaXiv.