notesum.ai

Published at November 1

Privacy Risks of Speculative Decoding in Large Language Models

cs.CL

cs.AI

cs.CR

cs.DC

cs.LG

Released Date: November 1, 2024

Authors: Jiankun Wei, Abdulrahman Abdulrazzag, Tianchen Zhang, Adel Muursepp, Gururaj Saileshwar

Arxiv: http://arxiv.org/abs/2411.01076v1

Refer to caption

	Accuracy (%)				F1 Score
TPQ $\rightarrow$	5	10	20	30	5	10	20	30
Experiment 1
LADE	53.6%	63.2%	65.5%	68%	0.49	0.59	0.63	0.66
REST	91.2%	97.6%	100%	99.6%	0.91	0.98	1.0	1.0
BiLD	28.4%	41.6%	49.2%	52%	0.27	0.43	0.52	0.55
Experiment 2
LADE	33.2%	53.2%	64.8%	70.8%	0.3	0.51	0.63	0.7
REST	96.4%	97.6%	100%	99.6%	0.96	0.98	1.0	1.0
BiLD	58.8%	68%	75.2%	77.6%	0.54	0.64	0.73	0.76
Experiment 3
LADE	15.2%	16%	13.6%	18.8%	0.13	0.13	0.1	0.16
REST	35.2%	38.4%	40%	39.6%	0.27	0.32	0.34	0.33
BiLD	13.2 %	20%	20.4%	20%	0.11	0.16	0.16	0.16