"Influence Sketching": Finding Influential Samples In Large-Scale Regressions

TL;DR

This paper introduces influence sketching, a scalable algorithm that identifies influential samples in large-scale regressions, improving data inspection for model robustness and vulnerability detection.

Contribution

The paper develops influence sketching, a novel scalable method embedding random projections into influence computation for large high-dimensional datasets.

Findings

Influence sketching accurately identifies influential samples in large datasets.

Removing influential samples significantly decreases model accuracy.

Influence sketching uncovers previously unidentified malware samples.

Abstract

There is an especially strong need in modern large-scale data analysis to prioritize samples for manual inspection. For example, the inspection could target important mislabeled samples or key vulnerabilities exploitable by an adversarial attack. In order to solve the "needle in the haystack" problem of which samples to inspect, we develop a new scalable version of Cook's distance, a classical statistical technique for identifying samples which unusually strongly impact the fit of a regression model (and its downstream predictions). In order to scale this technique up to very large and high-dimensional datasets, we introduce a new algorithm which we call "influence sketching." Influence sketching embeds random projections within the influence computation; in particular, the influence score is calculated using the randomly projected pseudo-dataset from the post-convergence Generalized Linear Model (GLM). We validate that influence sketching can reliably and successfully discover influential samples by applying the technique to a malware detection dataset of over 2 million executable files, each represented with almost 100,000 features. For example, we find that randomly deleting approximately 10% of training samples reduces predictive accuracy only slightly from 99.47% to 99.45%, whereas deleting the same number of samples with high influence sketch scores reduces predictive accuracy all the way down to 90.24%. Moreover, we find that influential samples are especially likely to be mislabeled. In the case study, we manually inspect the most influential samples, and find that influence sketching pointed us to new, previously unidentified pieces of malware.