Score ranks stealthy source manipulation of anomaly detectors
Original title: Quantifying and Prioritizing Stealthy Source-Manipulation Threats Against Multi-Source Anomaly Detectors
On a 17-channel telemetry benchmark, a detector blind to 64.7% of audited task and subset pairs still caught 82.8% of random single-channel corruption.
A paper proposes the task-conditioned visibility margin to score adversaries who control some inputs to a multi-source anomaly detector and aim at a specific downstream decision. The margin is a closed form, takes about 85 μs per certificate, and yields a minimum-evidence attack. On a 17-channel industrial telemetry benchmark, task-conditioned manipulations caused more than twice as many falsely reassuring errors as generic stealthy ones and more often evaded the alarm on four of five splits. Used for audit ranking, the margin found a median of four verified failures in ten checks, versus none for random order and two for alarm-based ranking, and it also ranked GDN, TranAD, and a five-source crop-yield panel.
- A detector flagged 82.8% of random single-channel corruptions.
- It was first-order blind on 64.7% of task and subset pairs.
- Audits covered one- to four-channel subsets.
- Median verified failures were four in ten checks, versus none and two for baselines.
From the original article
1 min read at arxiv.orgSecurity pipelines often fuse several data sources and protect a downstream model with a cross-source anomaly detector. Such detectors are usually judged by their clean false-positive rate and by how well they catch random corruption. Neither metric captures an adversary who controls some sources and targets a specific downstream decision.
By Indraveni Chebolu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.