Table 6. Comparison of different...

Table 6.

Comparison of different accuracies (%) by using defense methods where the conditions of imbalanced data and few shots are evaluated by using CIFAR-100-LT and CIFAR-FS, respectively. TRADES (Zhang et al., 2019; Wu et al., 2021) is supervised while our TRADES in section 3.5 is self-supervised

Methods	Imbalanced data			Few shots
Methods	AA	RA	SA	AA	RA	SA
RoCL*	15.94	18.09	44.86	36.64	43.04	58.13
ACL*	16.86	18.46	45.48	39.96	45.43	58.73
TRADES (Wu et al., 2021)	18.79	21.24	43.82	–	–	–
TRADES*	18.04	21.12	45.04	42.84	47.94	57.96
Our (TRADES)	19.14	22.18	48.74	45.61	49.34	59.34
RoBal (Wu et al., 2021)	18.80	19.59	50.49	–	–	–
FP (Dong et al., 2025)	–	–	–	46.29	49.90	59.75
Ours (AT, $B = 256$ ⁠)	21.06	23.40	49.89	46.52	51.51	60.84
Ours (AT, $B = 128$ ⁠)	20.43	22.96	50.32	47.03	52.20	62.70
Ours (AT, $B = 64$ ⁠)	19.12	20.84	50.12	47.81	50.42	59.94

Methods	Imbalanced data			Few shots
Methods	AA	RA	SA	AA	RA	SA
RoCL*	15.94	18.09	44.86	36.64	43.04	58.13
ACL*	16.86	18.46	45.48	39.96	45.43	58.73
TRADES (Wu et al., 2021)	18.79	21.24	43.82	–	–	–
TRADES*	18.04	21.12	45.04	42.84	47.94	57.96
Our (TRADES)	19.14	22.18	48.74	45.61	49.34	59.34
RoBal (Wu et al., 2021)	18.80	19.59	50.49	–	–	–
FP (Dong et al., 2025)	–	–	–	46.29	49.90	59.75
Ours (AT, $B = 256$ )	21.06	23.40	49.89	46.52	51.51	60.84
Ours (AT, $B = 128$ )	20.43	22.96	50.32	47.03	52.20	62.70
Ours (AT, $B = 64$ )	19.12	20.84	50.12	47.81	50.42	59.94

[ViewLarge]

Sharing Unavailable