Explanations can be manipulated and geometry is to blame¶
Dombrowski, A. (2019). Explanations can be manipulated and geometry is to blame. Advances in Neural Information Processing Systems 32.
Cited by¶
1 citation across 1 artifact.
Each citation links to the sentence it supports in the citing article.
Mechanisms¶
- Feature Ablation or Occlusion Test
- And like any attribution, an occlusion map can be run backwards — the cue and mask chosen after the fact to produce the saliency picture someone wanted.
This sourceShows that six post-hoc explanation methods—not occlusion maps—can be manipulated toward arbitrary target maps with hardly perceptible input perturbations while model output remains approximately constant.
- And like any attribution, an occlusion map can be run backwards — the cue and mask chosen after the fact to produce the saliency picture someone wanted.
Verification¶
This reference passed the adversarial substantiation pipeline: it was checked to exist and to support the claim it is attached to. See how references were verified.
Registry ID ref:0a8221391d07 · see in the full table