Search papers, labs, and topics across Lattice.
2
0
5
By bounding conditioning representation magnitude, DP-aware AdaLN-Zero tames heavy-tailed gradients in differentially private diffusion models, leading to improved performance under strict privacy budgets.
By strategically allocating differential privacy based on layer-specific vulnerability, LM-DP-SGD significantly improves the privacy-utility trade-off against membership inference attacks on intermediate representations.
