| { |
| "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." |
| } |
| ] |
| } |
|
|
