Skip to content

Privacy-Preserving Incentive Mechanism Design for Federated Cloud-Edge Learning

intro

  • cloud-edge computing
  • three-layer
    • edge devices, EDs
    • process data locally
    • acccess points, APs
    • nearby edge servers
    • upload to CC
    • cloud center, CC
    • further processing or aggregation
  • minimize overall latency
  • private data of ED entirely exposed to APs
  • federated learning (FL) scheme
  • EDs train models locally and only upload models to upper layers
  • preserve privacy while maintaining model accuracy
  • federated cloud-edge learning (FCEL) system
  • EDs give models to AP and aggregate → partial model; APs give partial models to CC and aggregate → global model
  • private data outputs may be attached by leveraging the sensitive information in these outputs
    • og data may be partially recovered (reverse engineer)
  • differential privacy, DP
  • upload model updates with noise pertubation for privacy
  • avoids high computaion & communication overhead
  • pertubations affect model → privacy-accuracy trade-off
  • add Gaussian noise permuation
    • noice scaler over a threashold → AP unable to revover data
    • large noice scale → hard to converge

goal & model

  • optimal contract design problem
  • obtain a global model with desired accuracy in a certain time period while preserving EDs' data privacy
  • APs motivate EDs to participate without knowing EDs' privacy senstivitiy
  • CC determine the monetary incentive for lower layers for max model accuracy, 3 layer Stackelberg game + optimal contract design problem
    • CC pay APs, compensate what APs pay EDs
    • APs pay EDs, compensate the data leakage of EDs
    • choose different privacy budget
    • contract theory
    • constraints
    • incentive compatibility, IC
    • individual rationality, IR
    • privacy budget \(\epsilon\) reasonable
    • total rewards < total incentive
    • 3 layer Stackelberg game, TLSG
    • CC 預期 APs & EDs 策略，先手
    • use gradient ascent to update coeff. until winthin threashold → optimized coeff.

results

  • as noise scale increase, data leakage decrease & test loss increase exponentially
  • APs ↑ \(U_{CC}\) ↓；EDs ↑ \(U_{CC}\) ↑
  • comparison with DP-FedAvg
  • DP-FedAvg
    • only has CC & EDs
  • PFCEL converges faster, achieves higher accuracy & lower test loss than DP-FedAvg
  • test accuracy & test loss are close th those with zero noise scales
  • EDs ↑ test accuracy ↑ test loss ↓
  • non-i.i.d. degree of the whole data set decreases