TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

safety

Adversarial Training Dropped the Input Gradients

A low-rank Householder expansion finds local perturbation directions from the forward pass.

Published Updated Story ID: mp-2026-08-29-017
Read the complete editionStory JSON

Summary

A low-rank Householder expansion finds local perturbation directions from the forward pass.

The reported regularizer cost the equivalent of 2.8 PGD steps per epoch and matched stronger iterative training only within limited relative perturbation budgets, falling away beyond the local regime.

Why it matters

A low-rank Householder expansion finds local perturbation directions from the forward pass.

Limits and context

  • The reported regularizer cost the equivalent of 2.8 PGD steps per epoch and matched stronger iterative training only within limited relative perturbation budgets, falling away beyond the local regime.

Key claims

  1. A low-rank Householder expansion finds local perturbation directions from the forward pass.

    Qualification: The reported regularizer cost the equivalent of 2.8 PGD steps per epoch and matched stronger iterative training only within limited relative perturbation budgets, falling away beyond the local regime.

    Evidence: source-2026-08-29-019

Sources

  1. arXiv preprint 2608.26963arXiv · primary research

Corrections

No corrections have been recorded for this story.