Table 2.

Comparison of the results for CL and different self-supervised AT methods. The methods in the bottom-half used the pseudo labels (PLs) (Fan et al., 2021)

CIFAR-10CIFAR-100
MethodsRASARASAMemory (MB)Time
SimCLR*0.0391.770.4866.866,4097.7 h
RoCL* (K=7)39.9378.3718.7949.5319,1012 d 9 h
ACL*44.2379.0420.9747.5111,6492d 2.1 h
DynACL*46.8279.1023.2147.2012,0232d 2.6 h
Ours (AT, K=3)48.7479.3426.4947.066,40920 h
Ours (TRADES)50.2177.4927.6248.968,6771 d 12.4 h
Ours (AT, B=512)51.1576.9027.2145.126,4091 d 4.6 h
Ours (AT, B=256)50.5576.3327.1045.085,9891 d 1.8 h
AdvCL (Fan et al., 2021)50.4580.8527.6748.3420,8753 d
AdvCL*50.2580.5827.1247.79
Ours (TRADES) + PL51.2980.8928.0548.4611,4072 d 9 h
Ours (AT) + PL52.5280.3628.0148.058,9852 d 0.6 h
Note(s):

*Indicates our results. The highest RAs (%) are italic. Memory and computationcosts are evaluated

or Create an Account

Close Modal
Close Modal