TADR Diffusion Module Lifts Anomaly Detector F1 on SMD When Data Blocks Are Missing

Industrial sensor systems regularly lose entire blocks of data. The network drops, equipment fails, and a "hole" appears across several channels and several time steps at once. The standard response is a two-step pipeline: first fill the gap, then look for anomalies. Researchers from Seoul have proposed Task-Aware Diffusion Refiner (TADR), a diffusion module that targets the detection task rather than reconstruction accuracy when it restores data. According to their results, TADR raises F1, a combined score of precision and recall for anomaly detection, relative to the unmodified detector on SMD (Server Machine Dataset, a set of server metrics). The measurement used the strict Point-to-Point (P2P) scheme, in which every timestamp is scored separately.

Why MSE-based imputation erases anomalies

A conventional imputer is trained on mean squared error (MSE) and on normal equipment operation. As a result, it fills a gap with the most likely normal pattern. If a block covers an anomaly, the partially observed deviation is smoothed back toward normal, and the detector never sees it.

There is also an opposite effect. On normal data, the boundaries of a reconstructed block produce artifacts that trigger false alarms. The authors call this imputation bias. They note that the Point Adjustment metric long hid this loss of precision by inflating recall regardless of the number of false positives. That is why they evaluate the method under strict P2P.

The paper also singles out security systems, such as sensors that detect prohibited substances. A rare event there may coincide with a data loss, and an imputer that fills the gap with normal values will destroy the very thing the system exists to catch.

How TADR works

TADR is a conditional diffusion model trained with the loss function of a frozen detector. The authors took the architecture and sampler from CSDI, a diffusion imputer that serves as the baseline. Only the training objective changes: the diffusion should sharpen anomaly peaks and preserve the normal background, rather than simply minimize reconstruction error.

The main engineering problem is training. Backpropagating gradients through the entire sampling chain is computationally infeasible. So the task-specific terms are applied to a one-step estimate of the clean window, which the noise prediction provides on its own. As a result, training costs no more than ordinary diffusion.

The detector stays frozen, so TADR plugs in without changing its structure. The authors note that the method works with detectors that score anomalousness by reconstruction error, and they consider Anomaly Transformer, TimesNet and KAN-AD, among others.

Where the method helps and where it does not

The evaluation ran on five seeds, with confidence intervals and paired significance tests. In the ablation, the anomaly-amplification objective itself contributes the most, and the iterative scheme adds a further gain over a single-pass refiner. The available text gives no numerical values for the gains, so readers should consult the paper itself.

The authors do not claim universal superiority. The gain grows with the share of continuous, mutually redundant channels and with the amount of anomaly evidence that the gap removed. On telemetry datasets dominated by one-hot command channels there is no effect, and there is none on PSM either. Adversarial manipulation of observed channels, data poisoning and concept drift were not part of the evaluation.

The load profile matters in practice. TADR was measured at batch size one on a single CPU core, not on server GPUs. With ten sampling steps, the module processes a window on one core and, thanks to its compact model, can serve many parallel streams at 1 Hz. That makes it suitable for edge devices. The code and per-seed results are published on GitHub. The takeaway for engineers: if gaps are real in your edge pipeline and your detector works on reconstruction, evaluate the imputer by the detection metric, not by MSE.