DeAI: decentralized artificial intelligence
  • Introduction
    • General Terminology
  • Landscape
    • Data Providers
    • Computing Power
    • Model Training Task
    • Challenges
  • Privacy Preservation
    • Data Process
    • Privacy Preserved Training
    • Federated Learning
    • Cryptographic Computation
      • Homomorphic encryption
      • Multi-Party Computation
      • Trusted Execution Environment
    • Challenges
  • Security
    • Data Poisoning
    • Model Poisoning
    • Sybil Attacks
    • Impact of Large Models
    • Responsibility
  • Incentive mechanisms
    • Problem Formulation
    • Contribution Evaluation
    • Copyright
  • Verification of Computation
    • Computation on Smart Contract
    • Zero-Knowledge Proof
    • Blockchain Audit
    • Consensus Protocol
  • Network Scalability
    • Local Updating
    • Cryptography Protocol
    • Distribution Topology
    • Compression
    • Parameter-Efficient Fine Tuning
  • Conclusion
Powered by GitBook
On this page
  1. Privacy Preservation

Challenges

Each privacy preservation approach has its drawbacks. Cryptographic methods guarantee a high level of privacy but suffer from significant efficiency drawbacks and may not defend against membership inference attacks. Differential privacy and adversarial regularization mitigate privacy attacks to some degree without introducing additional computation costs but may impact model performance. These challenges are often mutually incompatible. Furthermore, the strengths of these techniques can also be double-edged; cryptographic approaches prevent data leakage in communication but also hinder data auditing, potentially facilitating backdoor attacks and data poisoning. To mitigate these challenges, multiple techniques are often used together, with federated learning serving as a backbone to integrate with other techniques such as differential privacy, MPC, and cryptographic methods. Some studies introduce trusted third parties to address efficiency challenges.

Method
Model Performance
Efficiency
Network Requirement
Risk

DP

Lower

Similar

Low

FL

Similar

Sligntly slower

High

Gradient leakage

FHE

Much lower

Much slower

Low

MIA

TEE

Same

Much slower

Low

MIA

MPC

Same

Much slower

Very high

MIA

PreviousTrusted Execution EnvironmentNextSecurity

Last updated 1 year ago