About the attack success rate

[clee-jaist]

Hi, I did a little test today, and the parameter settings are the same as those in your previous code. I found that the poison test acc is basically above 90%. Does this refer to the attack success rate? I checked the evaluated code and found that the poison test and clean test codes are exactly the same, which makes me a little bit confused. I show the last 5 epochs of poison_train

Thank you so much for your time

................................................................................................... ===Epoch: 196/200=== Poison training...	-------------	---------	---------		Iteration	loss	acc
0/391	0.00037	1.00000
130/391	0.00073	1.00000
260/391	0.00076	1.00000
390/391	0.00817	1.00000

Training summary:	---------	---------	----------		loss	acc	time
0.00120	0.99996	11.87119

Test model on clean data...	-------------	---------	---------		Iteration	loss	acc
0/79	0.24480	0.96094
26/79	0.37069	0.91406
52/79	0.25858	0.91406
78/79	0.53093	0.93750
---------	---------	---------
loss	acc	time
---------	---------	---------
0.23299	0.94581	0.86842

Test model on poison data...	-------------	---------	---------		Iteration	loss	acc
0/71	0.08573	0.97656
23/71	0.03288	1.00000
46/71	0.03576	0.99219
69/71	0.02621	1.00000
---------	---------	---------
loss	acc	time
---------	---------	---------
0.04901	0.98834	0.78306

Adjust learning rate to 0.0010000000000000002 ===Epoch: 197/200=== Poison training...	-------------	---------	---------		Iteration	loss	acc
0/391	0.00092	1.00000
130/391	0.00142	1.00000
260/391	0.00096	1.00000
390/391	0.00048	1.00000

Training summary:	---------	---------	----------		loss	acc	time
0.00112	0.99994	12.01725

Test model on clean data...	-------------	---------	---------		Iteration	loss	acc
0/79	0.23334	0.96094
26/79	0.34683	0.91406
52/79	0.24993	0.91406
78/79	0.53390	0.93750
---------	---------	---------
loss	acc	time
---------	---------	---------
0.22779	0.94689	0.90150

Test model on poison data...	-------------	---------	---------		Iteration	loss	acc
0/71	0.12132	0.97656
23/71	0.06421	0.98438
46/71	0.06803	0.98438
69/71	0.05093	0.99219
---------	---------	---------
loss	acc	time
---------	---------	---------
0.07787	0.97810	0.81564

Adjust learning rate to 0.0010000000000000002 ===Epoch: 198/200=== Poison training...	-------------	---------	---------		Iteration	loss	acc
0/391	0.00123	1.00000
130/391	0.00086	1.00000
260/391	0.00099	1.00000
390/391	0.00146	1.00000

Training summary:	---------	---------	----------		loss	acc	time
0.00122	0.99994	11.99979

Test model on clean data...	-------------	---------	---------		Iteration	loss	acc
0/79	0.23626	0.96094
26/79	0.35506	0.92188
52/79	0.25306	0.91406
78/79	0.53051	0.93750
---------	---------	---------
loss	acc	time
---------	---------	---------
0.23139	0.94591	0.85511

Test model on poison data...	-------------	---------	---------		Iteration	loss	acc
0/71	0.08064	0.97656
23/71	0.03207	1.00000
46/71	0.03732	0.99219
69/71	0.02520	1.00000
---------	---------	---------
loss	acc	time
---------	---------	---------
0.04738	0.98900	0.79475

Adjust learning rate to 0.0010000000000000002 ===Epoch: 199/200=== Poison training...	-------------	---------	---------		Iteration	loss	acc
0/391	0.00082	1.00000
130/391	0.00040	1.00000
260/391	0.00080	1.00000
390/391	0.00155	1.00000

Training summary:	---------	---------	----------		loss	acc	time
0.00119	0.99998	12.02322

Test model on clean data...	-------------	---------	---------		Iteration	loss	acc
0/79	0.24507	0.96094
26/79	0.36788	0.91406
52/79	0.25810	0.90625
78/79	0.52386	0.93750
---------	---------	---------
loss	acc	time
---------	---------	---------
0.23248	0.94630	0.86541

Test model on poison data...	-------------	---------	---------		Iteration	loss	acc
0/71	0.11731	0.96875
23/71	0.06122	0.99219
46/71	0.07014	0.97656
69/71	0.04956	0.99219
---------	---------	---------
loss	acc	time
---------	---------	---------
0.07854	0.97777	0.80501

Adjust learning rate to 0.0010000000000000002 ===Epoch: 200/200=== Poison training...	-------------	---------	---------		Iteration	loss	acc
0/391	0.00054	1.00000
130/391	0.00046	1.00000
260/391	0.00050	1.00000
390/391	0.00155	1.00000

Training summary:	---------	---------	----------		loss	acc	time
0.00115	0.99992	12.06870

Test model on clean data...	-------------	---------	---------		Iteration	loss	acc
0/79	0.24703	0.96094
26/79	0.36921	0.91406
52/79	0.25322	0.90625
78/79	0.54743	0.93750
---------	---------	---------
loss	acc	time
---------	---------	---------
0.23348	0.94600	0.84732

Test model on poison data...	-------------	---------	---------		Iteration	loss	acc
0/71	0.12039	0.97656
23/71	0.06394	0.98438
46/71	0.07073	0.98438
69/71	0.05400	0.99219
---------	---------	---------
loss	acc	time
---------	---------	---------
0.07907	0.97755	0.81543

Adjust learning rate to 0.0010000000000000002

Process finished with exit code 0 .......................................................................................................

[hkunzhe]

1. The poison test acc is just the attack success rate.
2. The poison test data is wrapped with CleanLabelDataset. That is to say, triggers are added (no adversarial noise), and the labels remain consistent. Please refer to the code for details.

[clee-jaist]

1. The poison test acc is just the attack success rate.
2. The poison test data is wrapped with CleanLabelDataset. That is to say, triggers are added (no adversarial noise), and the labels remain consistent. Please refer to the code for details.

Thank you for your reply. I checked the details of the evaluation code, and found that you are right.