Datasets:
File size: 3,477 Bytes
6c1c833 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | {
"field_id": "F011",
"field": "adversarial machine learning",
"directions": [
{
"direction_id": "G0",
"name": "Adversarial input construction against deployed systems",
"scope": "Construct inputs that cause deployed prediction or generative systems to deviate from intended behavior. This includes optimization for individual inputs, cross-model transfer, target-model queries, feasible edits to structured inputs, physically realized visual attacks, and bypasses of generative-system safety mechanisms."
},
{
"direction_id": "G1",
"name": "Empirically robust predictor construction",
"scope": "Improve predictive performance under adversarial inputs through learning processes, representations, predictor structure, or model combination, encompassing the full research scope of each of these four mechanisms."
},
{
"direction_id": "G2",
"name": "Inference-time handling of adversarial inputs",
"scope": "Handle adversarial inputs at inference time by identifying attacks or affected regions, or by repairing or transforming inputs to preserve downstream predictions. Neither type of result requires the other."
},
{
"direction_id": "G3",
"name": "Provable prediction robustness",
"scope": "Establish provable guarantees for predictions under perturbation, or train models to satisfy such guarantees. This includes smoothing-based certificates and non-smoothing deterministic verification and certified training."
},
{
"direction_id": "G4",
"name": "Assessment and explanation of adversarial vulnerability",
"scope": "Determine when and why models are vulnerable to adversarial intervention, and assess whether reported attack or robustness claims are credible. This includes empirical evaluation of attack validity and explanations based on features, geometry, and training dynamics."
},
{
"direction_id": "G5",
"name": "Statistical foundations of robust learning",
"scope": "Study robust risk, learnability, generalization, and theoretical limitations imposed by hypothesis classes or learning objectives."
},
{
"direction_id": "G6",
"name": "Trigger-dependent model compromise and defense",
"scope": "Study associations between covert triggers and model behavior, including both implanting controlled behavior and discovering, cleansing, or repairing compromised models and examples. Work on both sides need not be conducted in the same paper."
},
{
"direction_id": "G7",
"name": "Non-triggered corruption of learning data",
"scope": "Manipulate learning outcomes through training examples, labels, or continuing learning feedback without relying on a preinstalled trigger activated at inference time."
},
{
"direction_id": "G8",
"name": "Adversarial protection of released content",
"scope": "Modify material before release, on behalf of its content holder, to impede another model’s unauthorized recognition, extraction, editing, imitation, or training on that material."
},
{
"direction_id": "G9",
"name": "Adversarial robustness of interactive decision-making",
"scope": "Study attacks and defenses for policies and bandit learners whose actions, observations, and feedback unfold through interaction over time. This includes single-agent and cooperative settings and robust policy training."
}
]
}
|