Robustness相关

Strong!

Posted by tianchen on November 24, 2020

Survey - A Survey of Privacy Attacks in Machine Learning

  • Privacy Preserving
    • NN model should be a black box with only input and output
    • attack could be at the model itself / acquire private training data from it
  • the threat model
    • 4 parts:
    • different acess levels:just have the API / have full access to the model / training setting
  • Attacks
    • black-box atatcks: usually atack pretrained model service from the cloud
    • white-box: have access to all model params / grads during training
    • partial whitebox: make strong assumptions to black box but no full access

    • attacks training / inference: attacks training means active attacker
  • Taxnomy of threats:
    1. Membership Inference: whether input x is used as part of training set
    2. Reconstrcut Attack: recreate samples from their label using the model
      • also called attribute inference / model inversion
    3. Property Inference Attacks: extract the dataset property which is not explicitly listed as label:
      • e.g. find the ratio of women/man in patient dataset when the gender is not the label
    4. Model Extraction: black-box, want to acquire model
      • sometimes wants to generate a substitiude model, replicate the decision boundary
      • often wants to be efficient, as few queries as possible
      • sometimes also predicting the attributes of the model, the hyper-pram of reg, the optimizer type, the activation type
  • Attacks against centralized supervised learning
    • membership attck(shadow training) - lies in that models behave differently when they see data does not belong to the trainig set
    • reconstruction attack
    • property inference attack
    • Model Extraction attacks:
      • view model param/hyper-parm in objective as the unknown
      • for linear binary classifier with d dimension only needs d+1 queries, for non-linear perceptrons use optimization techniques as BFGS/SGD
      • similar to Active Learning, which has an external oracle to provide label for inquery
      • some others view it as finding the most valuable data to query:
        • use data not synthectic but from other fomain
        • unsupervised techniques as MixMatch
  • Attacks against distributed learning
    • federated learning setting, its not safe cause each client could access the model parameters
  • Defenses
  • Differentiabel Privacy:
      • add gaussian / laplacian noise to the output as the \epsilon
  • other approaches:
    1. regularization
    2. prediction vector tampering(干预)
    3. model compression
    4. ensembel
    5. noisy data
    6. weight quant
    7. selective sharing
    8. membership inference