Backdoor accuracy can't reach 100%

[ybdai7]

I have some problems when i am trying to reproduce the experiment results in this paper.

I pull the source code and run it with params.yaml after setting is_poison=True and switching some report loss options to True. After finishing training i found that the test accuracy on backdoor tasks cant reach 100% as it is said in the paper, it usually remains to be around 70% sometimes lower. I am wondering if there is something need to be set particularly or something wrong with my params.yaml. Could someone plz help me?

here are my params.yaml and the final results.

type	image
lr	0.1
momentum	0.9
decay	0.0005
batch_size	64
no_models	10
epochs	10100
retrain_no_times	2
number_of_total_participants	100
sampling_dirichlet	True
dirichlet_alpha	0.9
eta	1
save_model	True
save_on_epochs	[10, 100, 500, 1000, 2000, 5000]
resumed_model	False
environment_name	ppdl_experiment_Jul.13_13.34
report_train_loss	True
report_test_loss	True
report_poison_loss	True
track_distance	False
track_clusters	False
modify_poison	False
poison_type	wall
poison_images_test	[330, 568, 3934, 12336, 30560]
poison_images	[30696, 33105, 33615, 33907, 36848, 40713, 41706]
poison_image_id	2775
poison_image_id_2	1605
poison_label_swap	2
size_of_secret_dataset	200
poisoning_per_batch	1
poison_test_repeat	1000
is_poison	True
baseline	False
random_compromise	False
noise_level	0.01
poison_epochs	[10000]
retrain_poison	15
scale_weights	100
poison_lr	0.05
poison_momentum	0.9
poison_decay	0.005
poison_step_lr	True
clamp_value	1.0
alpha_loss	1.0
number_of_adversaries	1
poisoned_number	2
results_json	False
s_norm	1000000
diff_privacy	False
fake_participants_load	False
fake_participants_file	data/reddit/updates_cifar.pt.tar
fake_participants_save	False
current_time	Jul.13_14.56.08
adversary_list	[0]

[ebagdasa]

Hey, I don't really support the code anymore, so it might not work well. Maybe try to test memorization first, i.e. mix test and train poison_images together.