Search papers, labs, and topics across Lattice.
School of Artificial Intelligence, Beihang University
2
0
3
4
By intelligently perturbing class prototypes based on their discriminative power, VPDR achieves a superior privacy-utility trade-off in federated learning compared to naive Gaussian noise.
Forget likelihood scores and fine-tuning tricks: LLMs leak pre-training data membership through subtle, yet detectable, deviations in their gradient updates.
