hypothesis
active
hypothesis:if-adversarial-attacks-exploit-interference-they-should-activate-more-features-than-clean-inputsIf adversarial attacks exploit interference, they should activate more features than clean inputs.
Prediction tested via L0.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Adversarial examples activate more features than clean inputs across all ResNet18 layers.associated_withsupportsFigure 6.
- Early layer effect.
- Moderate mid-layer effect.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Interpretation of Figure 3.
- Adversarially-chosen prompts induce denial-of-service attacks consuming 10× more compute than benign promptsfinding0.772Background finding illustrating the severity of uncontrolled reasoning slowdowns.
- Interpretation of layer4 L0 spike.
- Unsatisfying reduction.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.757L2 attacks max norm 0.1 of average input norm.
- Limitation of single-prompt selection.
- Robustness controls superposition.
- Applied security implication derived from the asymmetry finding.