concept-pack-first-light / concepts /layer4 /adversarialimage.json
Possum's picture
Upload concept pack first-light v1.0.1
95a9764 verified
Raw
History Blame Contribute Delete
2.6 kB
{
"term": "AdversarialImage",
"role": "concept",
"parent_concepts": [
"MultimodalSafetyRisk"
],
"layer": 4,
"domain": "SafetyAndSecurity",
"definition": "Images with carefully crafted perturbations designed to fool computer vision models while appearing normal to humans",
"definition_source": "SUMO",
"aliases": [
"AdversarialExample",
"AdversarialPerturbation",
"EvasionAttack"
],
"wordnet": {
"synsets": [],
"canonical_synset": "",
"lemmas": [],
"pos": "noun"
},
"relationships": {
"related": [
"ImageRecognition",
"ObjectDetection",
"RobustML"
],
"antonyms": [],
"has_part": [],
"part_of": []
},
"safety_tags": {
"risk_level": "medium",
"impacts": [
"evasion",
"misclassification",
"safety_critical_failure"
],
"treaty_relevant": false,
"harness_relevant": true
},
"training_hints": {
"positive_examples": [
"The adversarial image caused the classifier to misidentify a stop sign as a speed limit sign.",
"Imperceptible pixel perturbations can cause dramatic changes in model predictions.",
"Adversarial patches can be printed and placed in the physical world to fool detectors.",
"The FGSM attack generates adversarial examples using the gradient of the loss function.",
"Adversarial images exploit the gap between human and machine perception.",
"The adversarial perturbation was invisible to humans but completely fooled the model.",
"Adversarial images can cause misclassification in safety-critical computer vision systems.",
"The adversarial patch caused the object detector to completely miss the person.",
"Adversarial image attacks can transfer between different models trained on similar data.",
"The adversarial example exploited high-frequency features that humans don't perceive."
],
"negative_examples": [
"The image was unclear.",
"The model made an error.",
"The picture was confusing.",
"The classification was wrong.",
"The image was noisy.",
"An error occurred.",
"The model was confused.",
"The image was distorted.",
"The prediction was incorrect.",
"The model made a mistake."
],
"disambiguation": "Intentionally crafted perturbations to fool models, not natural errors or noise"
},
"is_category_lens": true,
"child_count": 0,
"meld_source": {
"meld_id": "org.hatcat/multimodal-safety@0.2.0",
"applied_at": "2025-12-10T20:54:17.844054Z",
"pack_version": "6.0.0"
}
}