Evidence · DiggingBeagle record
TMIA: temporal membership inference against continually updated models
November 4 paper measuring whether black-box observations of successive checkpoints reveal when training records were inserted or removed.
- Source role
- primary disclosure
Evidence record
November 4 paper measuring whether black-box observations of successive checkpoints reveal when training records were inserted or removed.
Claim-level citations (3)
- supportsModel update histories can reveal when training examples entered or left a model: TMIA infers temporal training-sample membership from a sequence of black-box checkpoint outputs; its reported Pythia-70M SST-5 timestamp-level F1 is 91.34% in the evaluated configuration.
Abstract and Pythia-70M SST-5 temporal evaluation tables
- supportsModel update histories can reveal when training examples entered or left a model: The paper finds materially weaker inference when only a portion of model checkpoints is observable; 25% checkpoint access yields 48.48% timestamp-level F1 in one experiment.
Sparse checkpoint availability ablation
- supportsModel update histories can reveal when training examples entered or left a model: The TMIA authors report timestamp/interval F1 of 72.82%/64.66% when 50% of model snapshots are observable, falling to 48.48%/36.47% with only 25% observable.
Section 7.3 Attacking with Partial Observation of Model Updates, Table 7