Two-stage training strategy

[PWRuiLin]

Can this use a two-stage training strategy?

[jzyustc]

Hi, this repo does not support two-stage training. However, I think it may be simple for you to add some codes to achieve two-stage training.

[PWRuiLin]

I am glad to receive your reply. I would like to know how you trained a model for single noise jpeg. Is it trained as follows? Is it convenient to disclose the specific details? {   "project_name": "MBRS",   "with_diffusion": false,   "only_decoder": false,   "epoch_number": 101,   "batch_size": 4,   "train_continue": false,   "train_continue_path": "/",   "train_continue_epoch": 0,   "dataset_path": "datasets/",   "save_images_number": 4,   "H": 256,   "W": 256,   "message_length": 256,   "lr": 1e-3,   "noise_layers": [     "Combined([JpegMask(50),Jpeg(50),Identity()])"   ] }   哈哈 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: ***@***.***>; 发送时间: 2024年1月12日(星期五) 下午5:24 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [jzyustc/MBRS] Two-stage training strategy (Issue #11) Hi, this repo does not support two-stage training. However, I think it may be simple for you to add some codes to achieve two-stage training. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[jzyustc]

Sure, we use the default settings in ./train_settings.json to train a model for JPEG attack.

[PWRuiLin]

Thank you, and I would also like to ask how to compare the robustness of single noise models with other methods. When training a certain single noise, do we also need to add real Jpeg and simulated JPEG.
Such as gaussian noise:
"noise_layers": [
"Combined([JpegMask(50),Jpeg(50),Identity(), GN((0.01)])"
]

[jzyustc]

When training for single noise, we did not add other noises. We only use simulated JPEG, real JPEG and Identity

[PWRuiLin]

May I ask, which types of noise and corresponding noise parameters were used in the training of the combined noise model in this experiment, as shown in the table?
Are these the ones in the following picture?

[jzyustc]

@PWRuiLin

For MBRS (w/o diffusion block), we use the same combined noises as in your picture.

For MBRS (w/ diffusion block), we find that using [Real JPEG, Simualted JPEG, Identity and Crop] brings better results. You can refer to Section 4.3.5 in the paper or #2 (comment) for more details.

[jzyustc]

Closing due to inactivity. Feel free to reopen if needed. Thanks.