Détection d’anomalies par erreur de reconstruction

Autoencoders are great at finding anomalies: data points that do not fit the normal pattern. The idea is clever and surprisingly simple.

You train the autoencoder on normal data alone. It becomes an expert at rebuilding the kinds of inputs it has seen many times.

For each input you measure the reconstruction error: how far the rebuild is from the original. This single number drives the whole method.

Familiar, normal inputs rebuild with low error. The model has learned their structure, so the output closely matches the input.

An anomaly looks nothing like training data, so the model rebuilds it poorly. That produces a high reconstruction error. 🚩

You set a threshold on the error. Anything above it is flagged as an anomaly; anything below counts as normal.

A common trick is to look at errors on a clean validation set and pick a high percentile, like the 95th, as your cutoff.

You never need labeled anomalies to train. That makes this approach perfect when bad examples are rare or unknown in advance.

This powers fraud detection, factory defect spotting, and network intrusion alerts, anywhere normal is common and problems are rare.

A low threshold catches more anomalies but raises false alarms. A high one stays quiet but may miss real problems. Tune for your cost.

If normal behavior changes over time, old errors mislead you. Retrain or recalibrate to handle this drift and keep alerts accurate.

How does an autoencoder flag an anomaly?

Train on normal data, measure each input's reconstruction error, and flag anything above your threshold. No anomaly labels required. 🎉