Text reconstruction from gradients on CoLA
100ROUGE-1DAGER
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DAGERModel=GPT-2_BASE, Batch size (B)=162024.05 | 100 | 100 | — | |
| DAGERModel=LLaMa-2 (7B), Batch size (B)=162024.05 | 100 | 100 | — | |
| DAGERModel=GPT-2_BASE, Batch size (B)=322024.05 | 100 | 100 | — | |
| DAGERModel=LLaMa-2 (7B), Batch size (B)=322024.05 | 100 | 100 | — | |
| DAGERModel=GPT-2_BASE, Batch size (B)=642024.05 | 100 | 92.9 | — | |
| DAGERModel=LLaMa-2 (7B), Batch size (B)=642024.05 | 99.9 | 99.9 | — | |
| DAGERModel=LLaMa-2 (7B), Batch size (B)=1282024.05 | 99.5 | 98.2 | — | |
| LAMPL2+L1Architecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.02 | 94.5 | 52.1 | 76 | |
| LAMPcosArchitecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.02 | 93.9 | 59.3 | 80.2 | |
| LAMPcosArchitecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.02 | 92 | 56 | 78.8 | |
| LAMPL2+L1Architecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.02 | 91.2 | 47.8 | 75.4 | |
| LAMPcosArchitecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.02 | 89.6 | 51.9 | 76.2 | |
| LAMPL2+L1Architecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.02 | 87.5 | 47.5 | 73.2 | |
| LAMPcosArchitecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.02 | 85.8 | 46.2 | 73.1 | |
| TAGArchitecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.02 | 82.9 | 14.6 | 55.5 | |
| DLGArchitecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.02 | 82.7 | 10.5 | 55.8 | |
| TAGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.02 | 78.9 | 10.2 | 53.3 | |
| DLGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.02 | 59.3 | 7.7 | 46.2 | |
| TAGArchitecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.02 | 43.9 | 3.8 | 37.4 | |
| LAMPL2+L1Architecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.02 | 40.3 | 9.3 | 35.2 | |
| TAGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.02 | 40.2 | 3.1 | 32.3 | |
| DLGArchitecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.02 | 37.7 | 3 | 33.7 | |
| DLGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.02 | 36.2 | 2 | 30.4 | |
| DAGERModel=GPT-2_BASE, Batch size (B)=1282024.05 | 30.3 | 13.7 | — |