DiSPaT
A tabular data anomaly detection framework based on self-play fine-tuning. It uses an LLM to serialize table rows into text and alternately optimizes a policy model and a discriminator under an f-divergence minimization framework, tightening the characterization of the normal distribution without requiring real anomaly labels.
Timeline (last 90 days)
DiSPaT is a method that strengthens a model's understanding of normal data through a self-play fine-tuning framework, for tabular anomaly detection.
DiSPaT is based on a theoretical framework of f-divergence minimization and derives a tight approximation, whose training objective is equivalent to narrowing the gap between the true normal data distribution and the distribution of samples generated by the model.
DiSPaT uses an alternating optimization mechanism: the policy model generates pseudo-anomalous samples, the discriminator learns to distinguish synthetic samples from real normal samples, and the discriminative signal is fed back to drive policy updates to align the distributions.
DiSPaT's training process does not require real anomaly labels; the model generates pseudo-anomalies itself and then learns to identify them, thereby continuously tightening its definition of normality.
DiSPaT shares ideas with generative adversarial networks, but its theoretical foundation is built on f-divergence minimization rather than JS divergence.
Experiments on diverse benchmarks show that DiSPaT consistently outperforms prior LLM-based methods, deep learning anomaly detection methods, and classic unsupervised detectors.
The computational cost of self-play training is higher than one-off fine-tuning, because each iteration must generate samples and train the discriminator.
The paper does not specify how pseudo-anomalies are generated, nor does it give the number of rounds and cost required for convergence.
For tabular-data-intensive scenarios such as industrial quality inspection, financial risk control, and log auditing, DiSPaT is expected to reduce reliance on anomaly annotations.
All Facts (9)
DiSPaT is a method that strengthens a model's understanding of normal data through a self-play fine-tuning framework, for tabular anomaly detection.
50%UnverifiedDiSPaT is based on a theoretical framework of f-divergence minimization and derives a tight approximation, whose training objective is equivalent to narrowing the gap between the true normal data distribution and the distribution of samples generated by the model.
50%UnverifiedDiSPaT uses an alternating optimization mechanism: the policy model generates pseudo-anomalous samples, the discriminator learns to distinguish synthetic samples from real normal samples, and the discriminative signal is fed back to drive policy updates to align the distributions.
50%UnverifiedDiSPaT's training process does not require real anomaly labels; the model generates pseudo-anomalies itself and then learns to identify them, thereby continuously tightening its definition of normality.
50%UnverifiedDiSPaT shares ideas with generative adversarial networks, but its theoretical foundation is built on f-divergence minimization rather than JS divergence.
50%UnverifiedExperiments on diverse benchmarks show that DiSPaT consistently outperforms prior LLM-based methods, deep learning anomaly detection methods, and classic unsupervised detectors.
50%UnverifiedThe computational cost of self-play training is higher than one-off fine-tuning, because each iteration must generate samples and train the discriminator.
50%UnverifiedThe paper does not specify how pseudo-anomalies are generated, nor does it give the number of rounds and cost required for convergence.
50%UnverifiedFor tabular-data-intensive scenarios such as industrial quality inspection, financial risk control, and log auditing, DiSPaT is expected to reduce reliance on anomaly annotations.
50%