finding
active
finding:adversarial-examples-activate-more-features-than-clean-inputs-across-all-resnet18-layersAdversarial examples activate more features than clean inputs across all ResNet18 layers.
Figure 6.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- If adversarial attacks exploit interference, they should activate more features than clean inputs.associated_withsupportsPrediction tested via L0.
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.
- Adversarial example transfer may be predicted by correlated arrangement of features in superpositionhypothesis0.764Speculative.
- Superposition maximalist stance.
- Interpretation of Figure 3.
- Limitation of single-prompt selection.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.743L2 attacks max norm 0.1 of average input norm.
- Future work.
- Models fully learn hacks.
- Out-of-distribution generalization of SAE features.