Intriguing Properties of Neural Networks

Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., & Fergus, R. (2014). Intriguing Properties of Neural Networks. International Conference on Learning Representations (ICLR 2014).

Type

Preprint

Intellectual base

Grey literature

Year

2014

DOI

10.48550/arxiv.1312.6199

Link

https://doi.org/10.48550/arXiv.1312.6199

Cited from

security_intelligence

Cited by

4 citations across 4 artifacts.

Each citation links to the sentence it supports in the citing article.

Primes

• Adversarial Boundary Navigation
  • Adversarial machine learning. Evasion attacks search a classifier's boundary for inputs that retain prohibited semantic content — still spam, still malware, still a stop sign — while crossing onto the allowed side.

    This sourceFirst demonstration of adversarial examples — minimal, often imperceptible perturbations that cross a classifier boundary while preserving semantic content.

    Supported in partVerified against the publisher's abstract

    Imperceptible perturbations flip an image classifier while preserving the object category — the evasion principle — but the source says nothing of spam, malware, or stop signs.

    “We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error.”

    Read by an automated reader; not a human review. How support was checked

Domain-specific

• Input Manipulation Attack
  • Adversarial examples in computer vision (Szegedy et al., Goodfellow et al.) add imperceptible pixel perturbations that flip confident image-classification labels

    This sourceSzegedy et al. support imperceptible perturbations that flip image-classification labels; the 'confident' qualifier comes from the co-cited Goodfellow et al., not this paper.

    Supported in partVerified against the work's full text

    “We term the so perturbed examples “adversarial examples”.”

    Read by an automated reader; not a human review. How support was checked

Mechanisms

• Adversarial Example Generation
  • The anchoring concept is adversarial examples

    This sourceDemonstrates deliberately optimized, visually imperceptible perturbations that cause trained neural networks to misclassify.

• Minimal-Pair Context Probe
  • It also maps the adversarial cue boundary

    This sourceDemonstrates that a small, adversarially selected perturbation can induce model misclassification.

Verification

Does it exist? Confirmed. This work's DOI resolves to a registered record, which fixes its identity. That is all it fixes.

Does it back the claim? Read against the text for 2 of 4 citations: 2 supported in part. Each verdict is shown under its citation below, with what in the work backs the sentence.

Was it audited? Yes. A second, independent pass read the citation against the article text and recorded a verdict.

Support is checked per citation rather than per work — the same source can be cited soundly in one article and wrongly in another. Per-citation recording began recently, so a citation with no recorded check is a gap in the record rather than evidence it went unchecked.

See how references were verified.

Links previously used in the corpus

Before the registry existed this work was also linked 2 other ways.

• https://arxiv.org/abs/1312.6199 ×1
• https://research.google/pubs/intriguing-properties-of-neural-networks/ ×1

Registry ID ref:67335b0d1db4 · see in the full table