Table 8.

The performance of the six attack methods on three datasets and the defense effect of ensemble distillation. BN denotes BadNets, RP denotes RIPPLe, Ins denotes InsertSent, HK denotes HiddenKiller, Style denotes StyleBkd

Dataset Victim BERT-base
BERT-large
Attack BN RP Ins HK Style SOS BN RP Ins HK Style SOS
SST-2 ASR 100 100 99.9 89.4 95.4 100 99.9 99.9 100 89.1 89.6 100
CA 91.6 89.1 90.8 90.4 91.7 89.7 92.9 92.0 92.9 92.8 93.2 93.2
△ASR 6.5 9.5 17.4 22.4 31.1 13.6 6.6 6.1 21.9 12.9 30.0 9.6
△CA 80.2 80.8 82.4 83.1 83.0 83.5 79.6 82.4 83.7 83.7 82.8 84.2
OLID2 ASR 100 99.9 99.9 90.3 81.4 100 99.9 99.9 100 90.5 81.0 100
CA 80.7 80.3 81.0 83.0 83.0 81.2 83.2 82.5 82.0 81.2 81.4 81.4
△ASR 10.4 10.8 19.2 27.9 28.0 13.9 10.0 9.4 21.7 16.9 25.1 11.0
△CA 73.1 75.5 74.0 73.6 73.0 72.8 71.3 73.1 72.1 75.1 74.0 72.8
AG News ASR 100 99.9 100 94.8 57.0 100 100 99.9 99.9 94.2 56.8 100
CA 94.2 92.0 93.0 93.8 94.6 93.2 93.3 93.8 93.7 94.3 94.6 94.5
△ASR 12.5 12.6 17.2 20.6 6.6 12.9 11.3 11.1 17.6 18.7 7.6 13.3
△CA 82.3 81.8 81.4 81.7 81.4 82.5 81.6 82.2 81.8 81.7 83.9 83.2

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.