Excessive Invariance Causes Adversarial Vulnerability

TL;DR

Deep neural networks suffer from adversarial vulnerabilities due to excessive invariance to task-relevant features, which can be mitigated by an information-theoretic loss extension that encourages models to utilize all relevant features.

Contribution

The paper identifies excessive invariance as a key cause of adversarial vulnerability and proposes an information-theoretic loss extension to address this issue.

Findings

Invariance occurs across various tasks and architectures.

Manipulating class-specific content can change model predictions without altering hidden activations.

An information-theoretic loss improves model reliance on all task-relevant features.

Abstract

Despite their impressive performance, deep neural networks exhibit striking failures on out-of-distribution inputs. One core idea of adversarial example research is to reveal neural network errors under such distribution shifts. We decompose these errors into two complementary sources: sensitivity and invariance. We show deep networks are not only too sensitive to task-irrelevant changes of their input, as is well-known from epsilon-adversarial examples, but are also too invariant to a wide range of task-relevant changes, thus making vast regions in input space vulnerable to adversarial attacks. We show such excessive invariance occurs across various tasks and architecture types. On MNIST and ImageNet one can manipulate the class-specific content of almost any image without changing the hidden activations. We identify an insufficiency of the standard cross-entropy loss as a reason for these failures. Further, we extend this objective based on an information-theoretic analysis so it encourages the model to consider all task-dependent features in its decision. This provides the first approach tailored explicitly to overcome excessive invariance and resulting vulnerabilities.