Mechanistic Anomaly Detection via Functional Attribution

Hugo Lyons Keenan, Christopher Leckie, Sarah Erfani · 2026

Read Mechanistic Anomaly Detection via Functional Attribution by Hugo Lyons Keenan, Christopher Leckie, Sarah Erfani (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-18970.txt.

Description

Keenan, Leckie, and Erfani reframe mechanistic anomaly detection as a functional attribution problem. Labels can check outputs, but cannot tell whether normal or anomalous mechanisms produced them. Their method asks how far trusted samples can explain an output, using influence functions and parameter space sampling. On vision backdoors it tops BackdoorBench, with an average Defense Effectiveness Rating of 0.93 across seven attacks and four datasets. It also beats baselines on language model backdoors, including obfuscated models, and detects adversarial and out of distribution samples.

Catalog record

CREATORHugo Lyons Keenan, Christopher Leckie, Sarah Erfani
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80699
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.18970

Browse the full library · About the Tower of Babel