Unverified50% confidenceFactExact time
The privacy amplification effect of DP-SGD relies on the premise that the batch size is not leaked.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedDP-SGD is a mechanism for achieving differential privacy in stochastic gradient descent, implemented by clipping per-sample gradients and then adding Gaussian noise.76% similarUnverified训练时应用差分隐私噪声（DP-SGD）可有效降低记忆化程度，但会带来个性化效果的损耗71% similarUnverifiedThe dummy-point mechanism addresses the privacy loss caused by batch size leakage and truncation, and does not change the noise calibration logic of DP-SGD itself.70% similarUnverifiedTraditional privacy-preserving alignment methods usually assume that the reward model parameters are dense, causing the upper bound on estimation error to grow polynomially with the model dimension d.66% similarUnverifiedFully local inference is significantly constrained in model scale by device computing power, requiring a trade-off between privacy protection and model capability.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1125571API
curl https://kongchang.com/api/v1/knowledge/claims/1125571MCP
get_claim(id=1125571)