KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Shared Secret
Cryptographic technique that divides a secret into multiple shares distributed among different participants, where no individual share reveals information about the secret, but a predefined threshold of shares allows its reconstruction.
Random Masking
Protection method consisting of adding random values to gradients before transmission, with a guaranteed global zero sum ensuring that the final aggregation remains unchanged while masking individual contributions.
Compliance Verification
Cryptographic mechanism allowing verification that gradients submitted by clients follow the expected protocol without revealing their exact content, thus detecting malicious behavior or transmission errors.
BGW Protocol
Secure multi-party computation protocol developed by Ben-Or, Goldwasser and Wigderson, allowing participants to collectively evaluate an arithmetic function on their private inputs without revealing them.
SPDZ Protocol
Advanced secure computation framework offering active security against malicious adversaries, combining secret sharing, MAC authentication and preprocessing to guarantee the integrity of distributed computations.
Differentially Private Aggregation
Approach integrating differential privacy into the aggregation process, adding controlled noise to aggregated gradients to ensure that individual contributions cannot be inferred from the final model.
Gaussian Noise
Random noise distribution following a normal law, used in differential privacy mechanisms to mask individual contributions while preserving the global statistical properties of aggregations.
Exponential Mechanism
Differential privacy algorithm that selects an output with a probability exponentially proportional to its utility score, offering an optimal trade-off between utility and privacy protection.
Yao's Protocol
Fundamental secure two-party computation protocol allowing two participants to evaluate a function on their private inputs without revealing them, using boolean circuits and cryptographic gates.
Secure Multi-Party Computation
Cryptographic paradigm allowing a group of participants to jointly compute a function on their private inputs without revealing these inputs beyond the function's result.
Cryptographic Hash Function
One-way mathematical function transforming an arbitrary-sized input into a fixed-size output, used to guarantee the integrity and authenticity of messages transmitted in aggregation protocols.
Digital Signature
Cryptographic mechanism based on asymmetric cryptography allowing authentication of the origin and integrity of transmitted gradients, ensuring that only an authorized entity can submit updates.
Zero-Knowledge Proof
Cryptographic protocol allowing a prover to demonstrate knowledge of information without revealing that information, used to verify the validity of calculations without compromising confidentiality.
Program Obfuscation
Technique transforming a program into a functionally equivalent but structurally complex version, making its analysis or reverse engineering difficult to protect aggregation algorithms.
Secure Weighted Aggregation
Variant of secure aggregation where client contributions are weighted according to metrics like data quality or computing power, while maintaining the confidentiality of individual weights.
Secure Gradient Compression
Technique combining gradient compression and cryptography to reduce necessary bandwidth while preserving confidentiality, often using secure quantization and entropy coding methods.
Accusation Protocol
Mechanism allowing participants to report abnormal or malicious behaviors during aggregation, with cryptographic proofs ensuring the authenticity of accusations without compromising confidentiality.
Differential Update
Approach where clients transmit only the differences between their current and previous gradients, reducing disclosed information and enabling more efficient aggregation with better privacy protection.