{
  "field_id": "F023",
  "field": "backdoor attacks",
  "directions": [
    {
      "direction_id": "G0",
      "name": "Trigger and poisoning design for backdoor activation",
      "scope": "Design samples or triggering conditions that enable effective backdoor implantation and activation while addressing constraints such as correct labels, trigger stealth and selectivity, and limited poisoning opportunities; these constraints need not all be present at once."
    },
    {
      "direction_id": "G1",
      "name": "Backdoor control of generated outputs and downstream uses",
      "scope": "Implant conditional malicious behavior to control model-generated content or to carry the behavior of compromised pretrained representations into downstream tasks, prompts, and uses; a single attack need not combine generated-output control with cross-task transfer."
    },
    {
      "direction_id": "G2",
      "name": "Backdoor implantation through model and deployment artifacts",
      "scope": "Establish conditional malicious behavior through model and deployment artifacts, including architecture, parameters or embeddings, compilation, or execution hardware, rather than using poisoned training examples as the controlling intervention."
    },
    {
      "direction_id": "G3",
      "name": "Backdoor forensics and contaminated-data cleansing",
      "scope": "Determine whether a backdoor exists, how it works, or which training samples are contaminated, and cleanse data when training samples are the objects of intervention; model diagnosis and data cleansing are parallel branches, neither a prerequisite for the other."
    },
    {
      "direction_id": "G4",
      "name": "Repair of infected models",
      "scope": "Intervene after the fact on a model that already has a backdoor to eliminate conditional malicious behavior while preserving normal capabilities as much as possible."
    },
    {
      "direction_id": "G5",
      "name": "Backdoor-resistant learning from untrustworthy data",
      "scope": "Alter learning and representation formation during training or tuning to prevent trigger associations in potentially contaminated data from becoming backdoors in task prediction, rather than repairing a fixed infected model after the fact."
    },
    {
      "direction_id": "G6",
      "name": "Analysis of backdoor efficacy and persistence",
      "scope": "Primarily explain or characterize how attack conditions, trigger choices, and subsequent model updates affect backdoor success and persistence."
    },
    {
      "direction_id": "G7",
      "name": "Triggered-input detection and test-time neutralization",
      "scope": "At inference time, identify whether the current input is triggered or remove possible trigger effects to prevent the deployed model from executing backdoor behavior."
    }
  ]
}