Anomaly Detection Methods

When labels are scarce or the attack is novel, you detect deviation from normal instead of similarity to known-bad. Taxonomy by how much supervision you have:

1. Unsupervised / one-class

• Isolation Forest (not deep — know it anyway): build random trees that split on random features at random thresholds; anomalies isolate in few splits (short path length) because they sit in sparse regions. Fast, robust, the standard tabular baseline; an interviewer will respect “I’d start with isolation forest as a baseline.”
• Autoencoder reconstruction: train encoder→bottleneck→decoder to reconstruct normal data; anomalies (off the learned manifold) reconstruct poorly → anomaly score = reconstruction error $|x - \hat x|$. VAE variant: use likelihood. Caveats: AEs can generalize to reconstruct anomalies too (too much capacity), and if training data is contaminated with anomalies they get learned as normal.
• Deep SVDD: train an encoder to map normal data inside a minimal hypersphere; score = distance from center. (Needs tricks to avoid the encoder collapsing everything to the center — no bias terms, fixed center.)
• Density/distance in embedding space: embed with any good encoder (contrastive or pretrained), score by kNN distance to training points — embarrassingly effective, easy to serve with ANN.

2. Sequence anomalies (the abuse-relevant one)

Model the event stream: train an autoregressive model (GRU or small causal Transformer, exactly the SASRec recipe) to predict each next event given history; anomaly score = perplexity (how surprised the model is by what this account actually did). Catches: bot-like mechanical regularity (inter-event time entropy too LOW — bots are anomalously predictable), scripted action sequences, sudden behavior change inside one account (possible takeover or sale — see AM-10 Adversarial Dynamics and Edge Cases). Score per-event and aggregate over windows; condition on account age/segment so “new user” isn’t auto-anomalous.

3. Weakly/semi-supervised middle ground

A few labels + many unlabeled: PU learning (Class Imbalance, Calibration, and PU Learning), label propagation over a similarity graph, anomaly scores as features for a small supervised head.

4. Practical truths to say in interviews

1. Anomalous ≠ abusive. Power users, speedrunners, QA testers are anomalies. Anomaly scores should route to investigation or feed a supervised layer, rarely auto-enforce.
2. Conditioning matters: score relative to the right peer group (cohort by age/region/platform) or everything new is “anomalous.”
3. Threshold = review capacity: set thresholds to fill the human-review budget, not at a magic score.
4. Anomaly detection is the novel-attack safety net around the supervised system — frame it as one layer of defense in depth.